SSC-351

AN INTRODUCTION TO
STRUCTURAL RELIABILITY
THEORY

.,
..1

Thisckcumcnthm ken a~vd

for public releaseandsalq itt
diahihtim MUnlimitul

SHIP STRUCTURE COMMITTEE

1990
 sHIP STRUCTUFF COMMITFF

The SHIP STRUCTURE COMMllTEE is constitutedto prosecutea r~earch programto improvethe hull
structuresof shipsand other marinestructuresby an extensionof knowledgepertainingto design,
materials,and methodsof construction.

RADM J. D, Sims, USCG, (Chairman) Mr. H. T. Hailer

Chief, Office of Marine Safety, S=urity AsscciateAdministratorfor Ship-
and EnvironmentalProtection buildingand Ship Operations
U.S. Coast Guard MaritimeAdministration

Mr. Alexander Malakhoff Mr. Thomas W. Allen

Director,StructuralIntegrii EngineeringOfficer(N7)
Subgroup(SEA 55Y) MilitarySea@ Command
Naval Sea SystemsCommand

Dr. Donald Liu CDR MiohaelK Parmele8, USCG.

SeniorVie President Semetary, Ship StructureCommittee
American Bureauof Shipping U.S. Coast Guard

Mr. WilliamJ. Siekierka Mr. Greg D. Woods

SEA55Y3 SEA55Y3
Naval Sea Systems Command Naval Sea SystemsCommand

The SHIP S~UCTURE SUBCOMMllTEE acts for the Ship StructureCommitteeon technicalmattersby
providingtechnicalcoordinaticmfor determinatingthe goatsand objdhws of the programand by
evaluatingand interpretingthe resultsin terms of structuraldesign,construction,and operation.

~ HIPPIN NAVAL SFA SYSTFMS COMMANQ

Mr. Stephen G. Arntson(Chairman) Mr. RobertA Sielski

Mr, John F. Cordon Mr. Charles L. Null
Mr. William Hanzalek Mr. W. Thomas Packard
Mr, PhilipG. Rynn Mr. Allen H. Engle

MHITARY
SEALIFT
COMMANR II s COAST GUARQ
Mr. AlbertJ. Attermeyer CART T. E. Thompson
Mr. Michael W. Touma CAPT DonaldS. Jensen
Mr. Jeffery E. Beach CDR Mark E. Nolt

MARITIME ADMINISTRATION

Mr. FrederickSeibold
Mr. Norman0. Hammer
Mr. Chao H, Lin
Dr. Walter M. Maclean

u.s C~AST GUARD ACAtlFMY

LT Bruce Mustain
Mr. AlexanderB. Stavovy
. . MERCHANT MARINE ACADFMY
NATIONAL CADFMY OF S E CES
Dr. C. B. Kim ~
L!. s, MVAL ACAD~ Mr. Stanley G. Stianaen

Dr. Ramswar Bhattachafyya SOCIEW OF NAVAL ARCHITECTS AND

Dr. William S~dberg

Dr. W. R. Porter
AMERICAN lR~ N AND-L lNSTITU~
wELDING RESEARCH COLINC~
Mr. AfexanderD. Wilson
Dr. Martin Prager
.,,

c2.-
 Member Agencies:
Address Correspondent to:

United States Coast Guad

Secretary, Ship Structure Committee
Naval Sea Systems &rnmand
U.S. Cuast Guard (G-m)
Maritime Administration
Amekan Bureau of Str@i-g
Ship 2100 Second Street S.W.
Washington,D.C. 20593-0001
Military
Sea/&&mtn& Structure PH:(2o2)207-0003
Committee FAX: (202) 257-0025 .

An Interagency Advisory Committee

Dedicated to the Improvementof Marine Sttuotures

SSC-351
December 3, 1990 SR-1310

AN INTRODUCTIONTO STRUCTURALRELIABILITY THEORY

AS structural designs and analysis techniques evolve t we find
that there is a need to more rationally assess uncertainties
associated with marine structures. Naval architects and
structural engineers recognize that it necessarY to 9ain a
greater understanding of structural reliability as it relates to
the safety and serviceability of marine structures. This report
provides an excellent introduction o tructural eli~~i~~t~
theory. It was developed as a tutorial and should prove
valuable reference document.

J%. SIPES
Rear Admiral, U.S. Coast Guard
Chairman, Ship Structure Committee
 Tmhnical Report Documentation Paaa
zr-

1. Report No. 2. Government Accession No. 3. Recipients Cgtolog No.

SSC-351
4. Title ond Subtitle 5. Rwport Dote
JANUARY
1989
ANINTRODUCTION
TO STRUCTURAL
RELIABILITYTHEORY 6. Performing Orgdniznti~n Code

8. PgtforminO Orgenizntion Rcpert No.

7. Author(s)
QM E. MANSOUR SR-131O
9. Performing Orgoniz~tion Name and Address 10. Work Unit No. (T RAIs)

MANSOUR ENGINEERING, INC

BERKELEY, CA 94708 11. Contract or G, Ent NO.
DTCG23-86-C-20054
13. Typoef Report and P~rind Covered
!
12. Sponsoring AgencY Nome ond Address

SHIP STRUCTURE
COMMITTEE FINAL
U.S. COASTGUARD
2100 SECONDSTREET,SW, 14. Sponsoring A~mcy Code

WASHINGTON,DC 20593 G-M

15. Supplementary Notes

This work was carried out under the sponsorship of the Ship Structure Committee
and its member agencies.
16. Abstruct

This report provides an ititroduction to the state-of-the-art in structural

reliability theory directed specifically toward the marine industry. Comprehensive
probabilistic models are described for the environment, wave loads acting on a
marine structure, its failure modes, reliability under extreme load, system
reliability, and fatigue reliability. Application examples of various models are
presented including the necessary information required to performsuch reliability
analyses. A description of a computer program for calculating the reliability
level of a marine structure is provided as an appendix.

17. Koy Words 18. Distriktion

Stat*m~tAVailable to the public froll
Reliability Theory National Technical Information Service
Probability Springfield, VA 22161 or
Reliability Models Marine Technical Information Facility
System Reliability National Maritime Research Center
Fatigue Reliability Kings Point, NY 10024-1699
19. SocuritY Clossif. (ofthizrmpert) ~. Security Clo8sif. (of this page) 21. No. of Pogms 22. Price

UNCLASSIFIED UNCLASSIFIED 354

---- ---- - .- -.
Form DOT F 17UU./ (8-72) Reproduction of completed page authorized

3 ,<,
L
 METfttCCONVHWONFACTORS
*

AmmdrnM9 ~lwM to MotrkMuwo9 Awoxima Gnw

wtwnYaJKrwm M
Swmlbd whmYou Kmw Muh*by To Fltd symbol

-Ilmws

cm
m MmMun
&m

d
d
d

19

TEMPE

2

t
cm
ABSTRACT i
ACKNOWLEDGEMENT ii
TABLE OF CONTENTS ...
Ill

NOTATION vi

1. INTRODUCTION AND SUMM&RY 1

1.1 Role of Reliability Analysis in a General Probabilistic
Design Procedure 1
1.2 Basic Concepts in Reliability 2
1.3 Necessary Information for Rdiability Analyis of
Marine Structures 6

2. LOAD INFORMATION REQUIRED FOR RELL4J31LITT

ANALYSIS OF MARINE STRUCTURES: 11
2.1 Probabilistic Representation of Environment - 11
2.2 Dynamic Loads and Response of a Floating Vessel
Considered as a Rigid Body 40
2.3 Long-Term Prediction of Wave Loads x)
2.4 Prediction of Extreme Wave Loads 51
2.5 Return Pe~ods of Extreme Events and Non-
Encouqter Probabilities
2.6 Stochastic Combination of Loads on a Marine ,..
Structure 64
3. STRENGTH INFORMATION REQUIRED FOR RELIABILITY
ANALYSIS OF MARINE STRUCTURES: 9)
3.1 Strength Variability and Modelling w
3.2 Limit States Associated with Marine Structures %
3.3 Analysis of Uncertainty 94
3.4 Random Error Analysis s
3.5 Uncertainties Associated with Ship Strength w

iii

s (j
{
4. BASIC RELIABILITY CONCEPTS BASED ON FULLY
PROBAB~ISTIC METHODS - LEVEL 3: 1(M
4.1 Introduction - Reliability Levels 108
4.2 The Basic Problem - Level 3 la
4.3 The Normal Tail and Margin of Safety 112
4.4 Probability of Failure of a Ship Hull Girder 113

5. LEVEL 2 RELIABILITY ANALYSIS: w

5.1 The Mean-Value First-Order Second-Moment
(MVFOSM) Method m
5.2 Improvements to the MVFOSM Reliability Index IZ
5.3 Correlated Random Variables 137
5.4 Trend of the Reliability Index for Eighteen Ships 13
5.5 Ship Safety Index Based on Non-linear Limit State
Function 147

6. LEVEL 1 RELIABILIIY ANALYSIS: 153

6.1 Derivation of Partial Safety Factms from
Level 2 Methods 153
6.2 Recently Developed Reliability-Based Codes 156

7. SIMULATION AND MONTE CARLO METHOD 170

7.1 General Concept 170
7.2 Use of Monte Carlo Method in Ship Structural
Reliability Analysis 171
7.3 Generation of Random Numbers For a Random
Variable With a Prescribed Continuous Probability
Distribution 172
7.4 Sample Size and Variance Reduction Techniques 175
7.5 Application Examples 176

8. SYSTEM RELIABILITY 1%
8.1 Introduction 1%
8.2 General Formulation 1%
8.3 Bounds on the Probability of Failure of a Series
System 198
 8.4 Bounds on the Probability of Failure of a Parallel
System
8.5 General Systems m

8.6 Probabilistic Network Evaluation Technique (PNET) m

8.7 Fault Tree and Event Tree Xn

8.8 Reliability Bounds for Ship Primary Strength 2)8

9. FATIGUE RELIABILITY 213

9.1 Introduction 213
9.2 Fatigue Analysis 216
9.3 Reliability Models m
9.4 Design Application - De@ils m
9.5 Fatigue Reliabili@ - System m

10. APPLICATIONS TO SHIPS AND MARINE STRUCTURES 259

10.1 Long and Short-Term Procedures 259
10.2 Application Examples m

H. CONCLUDING REMARKS AND RECOMMENDATIONS: 302

APPENDIX 1: HELPFUL INFORMATION m
All Weibull Distribution Parameters - Probability
Paper w
A1.2 The Safety Index Versus Probability of Failure
for Normal and Other Distributions 311
APPENDIX 2: COMIWTER PROGRAM CALREL FOR
PERFORMING RELIABILITY ANALYSIS 314

.,-

7 ~.,-
 NOT ATIO~

The following list defines the main symbols appeming in this report.

L= ship length
x= random variable
Px = mean of X
ax = standard deviation of X
fx(.) = probability density function of X
Fx(.) = distribution function of X
EX = mean square value of X
A = parameter of exponential distribution
l,k = parameters of Weibull distribution
co = wave frequency
H(u) = frequency response function
&(w) = wave spectrum
S@) = response spectrum
N,n = number of records or encounters
Zn = random variable representing extreme amplitude of
total bending moment in n-records
(pzJ.) = probability density function of ~
Pf = probability of failure
p~n=l = probability of failure for n = 1
[A1UA2] = union of two events Al and Az

ii = complementary event of Ai
IAlnAz] = intersection of the two events Al and A2
Cg = generalized cost
Ci = initial cost of construction plus maintenance
Cf = cost due to failure
M= safety margin
s= random variable representing strength
z ,= random variable representing total bending moment
g(.) = limit state function or performance function
Xi* = coor ates of most likely failure point in original space
%

vi
 Yi* = coordinates of most likely failure point in reduced space
Vx = coefficient of variation of a random variable X
0 = central safety factor
$( ),@{) = standard normal probability density and distribution
fuctions
correlation coefficient
direction cosine
safety index
partial safety factors associated with a random variable

Metric Cmwe.mien Table

lft = 0.3048 In
1 in. = 25.4 mm
1 psi = 6.894 kPa
1 ksi = 6.894MPa
1 lb-in. = 0.113 N-m
1ton-ft = 0.309 t-m

vii
L INTRODUCTION AND SUMMARY

Structural reliability theory is concerned with the rational treatment of

uncertainties associated with design of stmctures and with assessing the safety
and serviceability of these structures. The subject has grown rapidly in the last
decade as can be seen from the many recent books and proceedings published on
the subject [1,2,3,4]. It has evolved from a research topic to procedures and
methodologies of wide range of practical applications and has been used in code
development.

There is a need for naval architects and structural engineers to develop an

understanding of structural reliability theory and its application to marine
structure. The aim of the application is usually to achieve economy together with
an appropriate degree of safety. However, like other tools, stmctural reliability
theory can be misused if not well understood. It cannot be thought of as the
solution to all safety problems and it cannot be applied in a mechanical fashion.
There are also several shortcomings that must be clearly identified and
examined.

The objective of this work is to provide an introduction and summary of the

state-of-the-art in structural reliability theory directed SDecific aily towards the
marine industrv. To this grid, consideration is given to: (a) the kind and nature of
existing data on the design variables of a marine structure, and (b) the numerical
nature of the analvsiq of. complex structures that typically exist in the marine
environment. .,

1,1 Role of Reliability Analvsis in a General Probabilish c Desiw Procedure

In order to define the role of reliability analysis in a general probabilistic

procedure for the design of marine structures, Figure 1.1 is introduced.

Starting with a configuration of the marine structure and using random

ocean waves as input, the wave loads acting on the structure can be determined
(please refer to Figure 1.1). Generally, for primary design analysis the most
important loads are the large ones. Extrapolation procedures are usually used to
determine the characteristics of these large loads. In the case of ocean-going

... . .

11 L,v
vessels, for example, this is done either through the detemnination of a long-term
distribution of the wave loads [5,6,7] or through the evaluation of an extreme load
distribution [8,9,10,11] that may occur in a specific storm condition.

In general, wave loads acting on an ocean-going vessel include low-

frequency loads due to the motion of the vessel in waves as a rigid body. They also
include higher frequency loads (and response) due to slamming and springing
which can be determined by considering the ship as a flexible body. In principle,
these loads should be combined stochastically to determine the total wave load as,
for example, developed in [12,13].

Referring back to Figure 1.1, other loads beside wave loads occur on a
marine structure. These loads may be important in magnitude, though usually
less random in nature (except possibly for wind loads on offshore stmctures). For
example, in the case of ocean-going vessels, these loads consist mainly of
stillwater loads and thermal loads.

Following Figure 1.1, the response of the marine structure to the total
combined loads is determined and compared with the resistance or capability of
the structure. This comparison may be conducted through one of several
relia bilitv methods. Based on these methods, safety indices or probabilities of
failure are estimated and compared with acceptable ones. A new cycle may be
necessary if the estimated indices are below the acceptable ones.

1.2 Bas ic Conce~ tin Reliability

In order to illustrate some aspects of the procedure described in figure 1.1

and to introduce the basic concept in the reliability analysis, the following
example is given. Consider a simple beam subjected to a loading induced by the
environment, e.g. wave load. Traditionally, in the design of such a beam,
practitioners and designers have used fixed deterministic values for the load
acting on the beam and for its strength. In reality these values are not unique
values but rather have probability distributions that reflect many uncertainties in
the load and the strength of the beam. Structural reliability theory deals mainly
with the assessment of these uncertainties and the methods of quantifying and

2
1
m
E
rationally including them in the design process. The load and the strength are
thus modelled as random variables.

Figure 1.2 shows the probability density functions of the load and the
strength of the beam in terms of applied bending moment and ultimate moment
capacity of the beam, respectively. Both, the load Z and the strength S are
assumed in this example to follow the normal (Gaussian) probability distribution
with mean values 1.LZ=20,000ft-tin and ps=30,000 ft-ton, respectively, and standard
deviations of 0Z=2,500 ft-ton and a~=3,000 f&tan, respectively.

Figure 1.2. Load and Strength Probability Density Functions

We may now construct a simple function g(s,z), called the limit state
fuction, which desctibes the safety margin M between the strength of the beam
and the load acting on it, i.e.,

M = g(s,z) = s-z (1.1)

Both S and Z are random variables and may assume several values.
Therefore, the following events or conditions describe the possible states of the
beam,

4
 ,..
.,.,. :,..

(i) M = ~S,Z) <0 represents a failure state since this means

that the load Z exceeds the strength S.

(ii) M = g(S,Z) > 0 represents a safe state

(iii) M = g(s,z) = O represents the limit state surface (line in

this case) or the border line between the safe
and failure states.

The probability of failure implied in (i) above can be computed from t

pf= Pm=g(s,z)~o] = JJ fs~ (S,2)ds dz (1.2)

g(s,z) s o

where fs,z (s,z) is the joint probability density function of S and Z and the
domain of integration is over all values ofs and z where the margin M is not
positive, i.e., not in the safe state. If the applied load on the beam is statistically
independent from the beam strength the above equation can be simplified and
interpreted easily as:

pf = ;S(d f~(z)dz (1.3)

o

where Fs (.) and fz (.) are the cumulative distribution fhnction of S and the
probability density function of Z, respectively, both in this example, are Gaussian.

Equation (1.3) is the convolution integral with respect to z and can be

interpreted with reference to Figure 1.2. If Z=z, the conditional probability of
failure would be Fs(z). But since z<ZSz+dz is associated with probabili~ fz(z)dz,
integration of all values of z results in equation (1.3).

In our example, S and Z are both statistically independent and normally

distributed. Equation (1.3) can be thus shown to reduce to:

5
 Pf = @(-p) (1.4)

where O(O)is the standard normal cumulative distribution function and ~ is

called a safety index defined as:

(1.5)
=*

Notice that as the safety index 13increases the probability of failure pf as

given by (1.4) decreases. The safety of the beam as measured by the safety index ~
can be thus increased (see equation 1.5) by increasing the difference between the
means ~s-yz or decreasing the standard deviations as and az.

Substituting in equation (1.5) the numerical values for ys, pz, ~s and Cz
given in our simple beam example results in a safety index ~ = 2.56. Equation
(1.4) can be then used in conjunction with tables of standard normal cumulative
distribution function to yield a probability of failure= 5.23x 10-3.

~formation
1. N for Reliability Ana Ivsis of Ma rine % r ucture~

The preceding example and Figure 1.1 indicate that certain specific load
and strength information are necessary for performing reliability analysis of
marine structures. It is mostly in this area that reliability analysis of marine
structures differs from typical civil engineering structures. In this report
emphasis is placed on developing the required load and strength information for
marine structures.

Prior to estimating the loads acting on ships or marine structures a

statistical representation of the environment is necessary. This includes waves,
wind, ice, seismic and current. The last four items are more important for fixed
offshore structures than for floating vessels. The environmental information can
then be used as input to determine the loads acting on the structure. Typically, an
inputioutput spectral analysis procedure is used to determine the short-term
loads in a specific sea condition (stationary condition). The required transfer

6
function is determined from strip theory using the equations of motion of the
vessel or from a towing tank experiment. In offshore structures, Monsons
equation is usually used to determine the wave load trwsfer fiction.

Short term prediction of the loads is not sufficient for the reliability
analysis. Extreme values and long-term prediction of the maximum loads and
their statistics are more valuable. For this purpose order statistics and statistics
of extremes play a very important role. Gumbels theory of asymptotic
distributions is often used in this regard. In the long-temn prediction, the fatigue
loads, i.e., the cyclic repetitive loads which cause cumulative damage to the
structure must also be considered.

For complete description of this aspect of reliability analysis, methods of

combining the loads such as static and dynamic, including high and low
frequency loads, must be considered. In nature, many of these loads act
simultaneously, therefore, their combination must be evaluated for a meaningful
reliability analysis. The environment and load aspects are discussed in Chapter 2
of this report.

The second major component in the reliability analysis is the strength (or
resistance) of the marine stmcture and the evaluation of its modes of failure. In
this regard several limit ~tates may be defined such as the ultimate limit state,
fatigue limit state and seficeability limit state. The first is related to the
maximum load carrying capacity of the structure, the second to the damaging
effect of repeated loading and the third to criteria governing normal use and .
durability. Each of these limit states include several modes; for example, the
ultimate limit state includes excessive yielding (plastic mechanisms) and
instability (buckling failure).

Methods of analyzing uncertainties associated with the loads and the

strength of marine structures are impor~nt aspects of reliability analysis [14,15].
Generally, these uncertainties are quantified by coefficients of variation since in
most cases lack of data prevents the estimation of complete probability
distributions. Strength, modes of failure and uncertainty analysis of marine
structures are discussed in Chapter 3 of the report;

7
 Chapters 4,5 and 6 present three different levels of reliability analysis based
on the load and strength information discussed in Chapters 2 and 3. Examples of
application to ships are also provided in these chapters. Chapter 7 introduces
simulation and Monte Carlo techniques as a tool for use in the reliability analysis.
System reliability, which deals with redundancy of structures and multiplicity of
failure modes is discussed in Chapter 8. Chapter 9 describes a procedure for
fatigue reliability which requires separate treatment from reliability under
extreme load. Several application examples to ships and offshore structures are
given in Chapter 10. The last chapter of the report discusses some shortcomings
and offers concluding remarks. Appendices are given at the end of the report one
of which includes some helpfd information and another describes a computer
program for performing reliability analysis.

The reader of this report can get full benefit of the material presented if
he/she has a background (one course) in basic probability theory and statistics
including probability distributions, random variables, expectation of a random
variable, sampling theory, and estimation methods.

8
1.1 Madsen, H. O., Krenk, S. and Lind, N. C., Methods of Structural Safety,
Prentice-Hall, Inc., Englewood Cliffs, 1986.

1.2 Ang, A. H.-S., and Tang, W. H., Probability Concepts in Engineering

Planning and Design - Volume II. Decision, Risk and Reliability, John
Wiley and Sons, Inc., New York, 1984.

1.3 Theft-Christensen, P. and Baker, M., Structural Reliability Theory and Its
Application, Springer-Verlag, Berlin, 1982.

1.4 Konishi, I., Ang, A. H.-S. and Shinozuka, M. (editors), Structural Safety
and Reliability, Volumes I, 11 and 111, Proceedings of ICOSSAR, Kobe,
Japan, 1985.

1.5 Mansour, A., Methods of Computing the Probability of Failure Under

Extreme Values of Bending Moment, M.I.T Report No. 70-15, Sept. 1970;
also Journal of Ship Research, Vol. 16, No. 2, June 1972.

1.6 Lewis, E., Predicting Long-Term Distributions of Wave-Induced Bending

Moment on Ship Hulls, Spring Meeting, SNAME, 1967.

1.7 Mansour, A., Probabilistic Design Concepts in Ship Stmctural Safety and
Reliability, Transactions, SNAME, 1972.

1.8 Ochi, M. and Motter, E., Prediction of Extreme Values of Impact Pressure
Associated with Ship Slamming, Journal of Ship Research, Vol. 13, No. 2,
June 1969.

1.9 Mansour, A. and Faulkner, D., On Applying the Statistical Approach to

Extreme. Sea Loads and Ship Hull Strength, RINA Transactions, 1973.

Llo Ochi, M. K., Principles of Extreme Value Statistics and Their

Application, Proc. Extreme Loads Symposium., Oct. 1981.

9
1.11 Mansour, A. and Lozow, J., Stochastic Theory of Marine Vehicles
Slamming Response in Random Seas, Journal of Ship Research, Vol. 26,
No. 4, Dec. 1982.

1.12 Femo, G. and Mansour, A., Probabilistic Analysis of the Combined

Slamming and Wave-Induced Responses, Journal of Ship Research, Vol.
29, No. 3, Sept. 1985, p.170.

1.13 Mansour, A., Combining Extreme Environmental Loads for Reliability-

based Designs, Proceedings of the Extreme Loads Response Symposium,
SSC and SNAME, Arlington, VA, October 1981, pp. 63-74.

1.14 Faulkner, D. and Sadden, J. A., Toward a Unified Approach to Ship

Structural Safety, Trans. Royal Institution of Naval Architects, Vol. 121,
1979, pp. 1-38.

1.15 Guesdes Soares, C. and Moan, T., Statistical Analysis of Stillwater Load
Effects in Ship Stmctures, Trans. SNAME, Vol, 96, 1988.

10
2. LOAD INFORMATION REQUIRED IN RELIABILrrY ANALYSIS OF
MARINE STRUCTURES

2.1 Probabilistic Representation of Environment

Prior to discussing the loads acting on a ship or a marine structure, a

discussion of a probabilistic representation of the environment is essential.
Information on the environment can be then used as input to determine the
loads acting on the structure. A complete description of the environment
entails description of waves, wind, ice, seismic and current. The last four
items are more important for fixed offshore structures than for floating
vessels. Since the main emphasis in this work is floating vessels, only waves
will be thoroughly investigated in this report.

The sea surface is irregular and never repeats itself. An exact

mathematical representation of it as function of time, wind speed, wind
direction, current, etc. is not possible. A representation, however, using a
probabilistic model is possible and more suitable. By means of the theory of
random processes one may represent the sea surface and determine certain
statistical averages and extreme values suitable for design.

Such a probabilistic representation of a random phenomenon has been

well developed in electrical engineering to analyze random noise (see
reference [2.1] ) and was used successfully in mechanical engineering to
investigate random vibration (e.g. [2.2]). It has been also used in civil
engineering for earthquake analysis. In the next section a few definitions
related to random processes and the associated probability distributions are
discussed.

2.1,1. Definitions:

Deterministic process: If an experiment is performed many times

under identical conditions and the records obtained are always alike,
the process is said to be deterministic. For example, sinusoidal or

11

2/
predominantly sinusoidal time history of a measured quantity are
records of a deterministic process [2.2].

Random process: If the experiment is performed many times when

all conditions under the control of the experimenter are kept the
same, but the records continually differ from each other, the process
is said to be random [2.2].

The degree of randomness in a process depends on: (1) the

understanding of effects of the factors involved in the experiment,
and (2) the ability to control them.

As an example of a random process, consider a test is being

performed to determine the wave elevation as a function of time at a
given location in the ocean. Figure 2.1 is a record of the wave
elevation as recorded for a period of approximately 18 minutes.

RHVE ELEVRTION RECORDS

0 Icc 260 300 400 500 6M 700 800 500 1Goo

TIIIE[SeC]

Figure 2.1

The same test was repeated under identical conditions as far as

is known, that is, under the same wind speed and a record
as shown in Figure 2.2 was obtained. The striking feature is that the
two records are not identical. If the test is repeated several times
under identical conditions as far as is known, records will be
obtained that are not identical. This randomness in the records is

12

22
due to factors beyond ones control and ability to measure.
The
elevation of the water surface at any time is due to the entire history
of the meteorological conditions in that area and surrounding areas.
Therefore, under given macroscopic parameters such as wind
direction, speed, duration, etc., one cannot predict exactly the wave
surface at the given point. The wave elevation records can be thus
treated as records of a random process.

.+ 1
[
o 100 2CC 300 400 500 600 700 800 930 1000
TIME[SeC]

Figure 2.2

Another example of a random process is an ensemble of time

history recordi of a - strain gage installed in a ferry boat operating
between, say, San FTancisco and San Rafael. In any given day, a
record during each *trip between these two neighboring cities is
obtained. The resulting ensemble of records can be treated as ~
records of a random process.

As discussed above, the most important notion involved in the

concept of a random process X(i)(t) is that not just one time history is
described, but the whole family or ensemble of possible time
histories which might have. been the outcome of the same test are
described. In the example of recording wave elevation at a given
point in the ocean, the end result is an ensemble of records of wave
variation as a function of time (see F&gum a. 3) .

13

23
 Each of the above records is called a sample. Some of these
samples are more probable than others, and in order to describe the
random process further it is necessary to give probability
information.

It should be noted that a sinusoid in a deterministic process

can be characterized completely by its amplitude and frequency
(phase is unimportant in many cases). Similarly, the random
samples can be characterized by some average
amplitude (root mean square) and a decomposition in frequency
(spectral density) as will be discussed later.

WRVE ELEVFITION RECORDS

4
2

-2

-4
0

x(z) (t)

4 .
2

x(n) (t) O
-2
-1
o

TIME[SHC]

Figure 2.3.
 A designer given the record ensemble shown in Figure 2.3 may:

a. Select the largest value in the ensemble and use it for his
new design with a factor of safety. In this case he will make no use
of all information he is given except for one value, i.e., the maximum
value, Orl

b. Try to obtain statistical information from all the records

and use such information in his new design.

If b is selected, a probability description of the random

process is necessary.

First and Second Order Probability Distrib utions:

At a fixed value, t = tl, (see Figure 2.3) the values. of X(i)(t 1),
which represent wave elevation, can be described by a graph such as
in Figure 2.4. This graph shows the probability density functionl
(p.d.f.) f [ x (tl ) ] or simply f(x) which has the following properties:

F(%)

PROBABILITY DENS m FUNCTION

Figure 2.4.

1 First order probability distribution

l-s
1) The fraction of the ensemble members for which the wave
elevation X, treated as a random variable, lies between x and
X + dx is f(x) dx I i-e. ,

P[x<X~x+&]=f(x)dx

2) The probability that a sample of wave elevation lies between a

and b is
b

p[a<X~b] = f(x)dx
f
a

3) The probability that X lies between - ~~ and + = is one, that is,

rm
p[.co<x~~] =
r f(x) dx = 1
m
4) P[x= a]= O.
*= a is a mnstant.

If, at two time instants t 1 and t2, the wave elevations treated
as random variables are denoted X(t 1) and X(t2 ), or simply XI and
X 2, the probability distribution of these (called joint probability
density function) is given by a surface f(xl ,x2) and has the
properties (Figure 2.5):

1) P[xl < Xl s Xl+ dxl and x2 < X2 <x2 + dx2] = f(xl,x2) dxldx2

~~

2) P[al <Xl ~ bl and a2 e X2 s b2] =

W *
u Z[

az al
f(xl ,x2) dxldx2

3) P~m< xl <++ -<X2< +to] =

/1
-m -@
f(xl ,x2) dxldx2 = 1

2 Second order probability distribution.

16
Ensemble Averages:

For a given function g(x), it may be defined:

E[g(x)] =expected value of g(x)= g(x)f(x)dx

1

.
,.
x,

* .

OENSl T~ FuNCTION
JOINT PQOOAW--l T+

Figure 2.5.

17
 For a given p.d.f. f(x), the following definitions will be used for
the wave elevation treated as a random variable X. When g(x) is
simply x, then
+m
E[x] =
--
J xfxdx
() = mean or ensemble average (2.1)
= expected value of X (analogous to
moment about origin or distance from origin to center of mass of a
rod of unit mass)

When g(x) = x2, then

E[x2] = j- x2f ( x )d x = mean square of the random variable X (2.2)

(analogous to m%ment of inertia or square of radius of gyration about
origin)

Root mean square (r.m.s.) = ,W

Setting g(x) = (x - E[x])2

+-m
~2 = E(x - E[x])2 = (x-E[x])2f(x)dx (2.3)
J
-e
= E[x2] - (E[x])2
= variance of the random variable X
(analogous to moment of inertia or square of radius of gyration about
center of mass)

r= ~2 = standard deviation of the random variable X.

/
At two fixed values tl and t2, let xl and x2 denote x(tl ) and
x(t2) respectively, then the autocorrelation and covariance functions
are defined by:
+= +-
Autocorrelation = E[xl ,x2] = xlx2f(xl ,x2)dxldx2 (2.4)
[[
(analogous to a product of inertia of a plate of unit mass about origin)

Covariance ~, = E { [xl - EIx1] ] [x2 - E[x211 }

xi X2 += +*
= [xl - EIx1]] [x2 - E[x2]] f(xl ,x2)dxldx2
H
-w -w

=
E[x1x2] - E[x11E[x21 (2.5)

(analogous to the product of inertia of a plate of unit mass about the

center of mass)

It should be noted that the covariance is equal to the

autocorrelation minus the product of the means. Therefore, if one of
the means is zero then the covariance is equal to the autocorrelation.
The correlation coefficient &l ~z can be defined as:

(2.6)

that is, a non-dimensional covariance.

The two random variables Xl and X2 are said to be

independent if:

f(x,, XL) = f(xi) f(x) (2.7)

therefore, from the definition of the autocorrelation function, it is

easy to show, in this case, that

E [x,xL]= E [x,] E [X2] (2.8)

19

27
 and, thus, both the correlation coefficient and the covariance
A, $~
~i, ~z are zero. This means that independent random variables must
necessarily be also uncorrelated (h mnmrse is no~ necessarily true) .

Note that when t ~= t, the covariance becomes identical with the

variance and the autocorrelation becomes identical with the mean
square.

It is interesting to notice that in rigid body dynamics we need

to know only the gross moments of inertia, but in vibration analysis
more detailed information on the inertia distribution is necessary.
Similarly, the average quantities derived above (mean,
autocorrelation, covariance, etc.) give only gross statistical estimates
of a random process. More refined estimates require more detailed
information about the probability distribution.

2.1.2 Stationary and Er~odic Processes:

A random process is an infinite ensemble of sample functions.

In the foregoing, the properties of the random process representing
the wave generation at fixed instants t ;, ,tz, etc., have been examined.
Next, the variation of these properties when t,, t ~are assumed to vary
is briefly discussed.

A random process is said to be stationary if its distributions are

invariant under a shift of time scale, that is, independent of the
origin.

This implies that the first order p.d.f. f(x) becomes a universal
distribution independent of time and E(x) and r k are also constants
independent of time. .

In addition, the second order p,d.f. is invariant under

translation of time scale; therefore, it must be a function of the lag
between t , and t ~ only, and not t ~ and t ~ individually. This implies

20
that the autocorrelation function is also a function of ~= tz - tl only
(see definition of the autocorrelation function given by equation 2.4).

E[x, x:]= E[x(t)x(t+?)]= R(?) (2.9)

where R( ? ) will be used to denote the autocomelation function of a

stationary random process. Note that R(0) = E[xz(t)] = mean square of ~
the process.

R( ~ ) is an important function because it comelates the random

process at any instant of time with its past (or future). R ( & ) has the
following properties (see figure 2.6):

i- R(0) = E[x J ] = mean square of the process

ii - R(+?) = R{- E) ie, an even function of t
...
111 - R(0) 2 pw)l R(&)
T
= mean square

z-

\
4
Figure 2.6. Autocorrelation function

If changes in the statistical properties of a random process

occur slowly with time then it is possible to subdivide the process
into several processes of shorter duration, each of which may be
considered stationary. It is usual to represent ocean waves as a
stationary random process over a period of about 30 minutes to two
hours .

The er~odic hypothesis states that a single sample function is

quite typical of all other sample functions; therefore, we can estimate

21
various statistics of interest by averaging over time instead of
averaging over the ensemble.
f(t)

Figure 2.7. A sample function f(t).

If f(t) represents such a sample function (Fig.2.7), then the

following temporal averages can be determined.

The temporal mean is,

<f> = limit 1 +!c/2 f (t) dt

~ [ (2.10)
T+- -T/2

The temporal autocorrelation function @ ( & ) is,

$(~) = limit + /+T/2 f(t)f(t + ~)dt

T+rn -T/2 (2.11)

The temporal mean square is,

<f2> = limit ; /+T2f2 (t) dt

T+= -T/2 (2.12)

An ergodic process implies that E[x] =< f>and R(F) =#( &), An
ergodic process is necessarily stationary since < f } is a constant while
E[x] is generally a function of the time t = t ~at which the ensemble
average is performed except in the case of stationary process.
However, a random process can be stationary without being ergodic.

22

32
 For ocean waves, it may be necessary to assume the ergodic
hypothesis if there is only one sample function of the process.

2.1.3 Spectral Density of a Stationary Random Process:

In many engineering problems it is customary to conduct a

Fourier analysis of periodic functions. This simplifies the problem,
because linearity permits treating each single frequency Fourier
component separately and finally combining to obtain the total
response.

A frequency decomposition of the autocorrelation function

R( T ) of the ocean waves can be made:

R(T). = ~+m S (u) eiudu -m <?<@

-m (2.13)

where S ( UJ) is essentially the Fourier transform of R{ ? ) (except for

the factor 2 IT ) given by

.
s(u) = *, f+m R(~)e-iwTd~ - JJ< m
.- (2.14)
* .

Relations (2.13) and (2.14) are known as Wiener-Khintchine ~

relations. It can be shown that S( L ) is a non-negative even function
of w [2.2]. A physical meaning can be given to S( OJ ) by considering
the limiting case when P=(),
krc
R(0) = mean square = [Ex&]= S(ti)d J ,
J
-m
that is, the mean square of the process = the sum over all frequencies
of S(ti)dti , so that S( * ) can be interpreted as a mean square spectral
density.

23

3?
 The mean square (area under the spectral density curve) is the
average of the square of the wave elevation and the root mean
square (rms) is the square root of that value. Physically, the larger
the mean square (or the r.m.s. value), the larger is the wave
elevation and the higher is the sea state.

Since the spectral density is an important function in ocean

waves, the following remarks are made:

1) The units of ocean waves spectral density are [ft%-see] since

+=

R (0) = E[x2] =
J
,
S(w ) dti; [ftk] ; therefore

s(u) = [ft* ]/units of circular frequency = [ft -see]

2) From the properties of a Fourier transform, it can be shown

that S ( u ) is a real and even function ofo. It can be represented by a
curve as shown in Figure 2.8.

w +(,d

Figure 2.8. Two-sided Spectral Density.

3) In practice, the negative frequency obviously has no

significance. It appeared in the mathematical model only to make
the sums easier. Because of the shape of S( Q ), it is called a two-sided
spectrum.

24
4) In practice, a one-sided spectrum can be defined by simply
folding the S( @ ) cume about the w = O axis, that is,
m
J S(L )dw = ~2S(~)dU
a square of the p;ocess, where
= ~~ti
o
)dti = mean

S(U) = one-sided spectrum = 2s(ti) ti~o

= o (d<o

S+(w ) is shown in Figure 2.9.

z, I

Figure 2.9. Energy Spectrum.

5) It can be shown that the area under the ocean waves spectral
density, that is, the mean square is proportional to the total energy
per unit area of the waves surface which is given by:

Total energy per unit area of the waves surface =

For this reason S+( u ) is sometimes called the energy spectrum.

25
 The energy in an increment Sw of wave frequencies at a cenbal
frequency u. is

Pg
(2.15)

6) From simple gravity waves, the total energy which is composed

of half kinetic energy and half potential energy is,

Energy per unit area = ~~$ $:, where $a= wave amplitude.

It follows that the square of the amplitude of a wave having

the same energy as all the wave components in the band of
frequencies represented by SW is:s

therefore,

~: = 2s(ti,)$w = double the incremental area under the S+( w )

curve

7) For this reason, oceanographers define a new eal density.- calkd.

the amplitude ~-~1 asiw obtained by doubling the S+ ( u ) ordinates.
The incremental area will then represent component wave
amplitudes as SW+O. The area under the amplitude spectrum
m
= 2 j S+(4) dti= 2E[x2]
G
= 2 mean square of the process. (2.16)

In tiis re~rt the energyrather than the anplitude spectral density will
k used.

3 Valid only for the limiting case when SW + ().

-<
26

?6
 It should be noted that, both the spectral density and the
autocorrelation function are measurable quantities that can be
indeed determined from time history records of ocean waves.

2.1.4 NarTow and Wide Band Random Processes:

A random process is said to be a narrow-band process if S( u )

has significant values only in a band or range of frequencies whose
width is small compared with the ma nitude of the center frequency
f
of the band UP . Figure 2.10 shows S ( U ) of a narrow band process
and the corresponding time history of a sample function. It should
be noted that a narrow band of frequencies appears in the sample
and it is meaningful to speak of individual cycles.

Qb w

Figure 2.10. Spectrum of a NarTow Band Process

(The sample shows a narrow band of frequencies.)

A random process is said to be a wide-band process if S( @ ) has

significant values over a band range of frequencies which is roughly
the same order of magnitude as the center frequency LJa. Figure 2.11
shows a typical S+ ( u ) and the corresponding sample function of a
wide-band process. Notice that a wide range of frequencies appears
in the sample function.

27
 Figure 2.11. Spectrum of a Wide Band Process,
(A broad range of frequencies are shown in the sample.)

2.1.5 Additional Statistics of a Random Process:

The rate of crossing a certain level of wave height or, in

general, a threshold is an important information in design. Similarly,
the probability distribution of the wave peaks can be useful in
estimating probabilities of exceedence of specified wave heights in a
given sea state. Because of their importance the rate of crossing a
threshold and the peak distribution of a random process will be
discussed in the next two sections:

a. Rate of Crossing a Threshold:

The problem of crossing a threshold was examined extensively

by Rice [2. 1]. Some of the important results of his work will be given
here without proof. Consider a random process #( t ) representing
wave heights and the process has a zero mean, i.e., E [x] = O. The
mean rate of crossing a given level a denoted by M; with positive
slope, that is, from below, (see figure 2.12) was derived by Rice [2.1]
as:

(2.17)

28
 x(t)

Figure 2.12. Crossing a Level a.

where & = d%
~d f%,~(., ., . ) is the j.p.d.f. of z and ~
Jk J
Similarly, the mean rate of crossing, that is, the average
number of crossing per unit time with a negative slope (from above)
is .
0
v;(t) = /11 ~ fX, ~ (a, i,t)di
. -m (2.18)
. .
If the threshold level a is zero, the corresponding mean rate of
crossing (from above and below) is

(2.19)

If the process is stationary- and narrow band, then Y: or Y; is the

apparent (mean) frequency of the process and from the stationarity
property they are constant, i.e., independent of time.

29
b. Peak distribution of a stationary narrow band random process:

For a narrow band random process (e.g. ocean waves) every

zero crossing from below is followed by a positive peak (crest), and
every zero crossing from above is followed by a negative peak
(trough). Therefore, the proportion of the positive zero crossings that
also cross the level a with a positive slope represents the
probability that the positive peak is larger than a, that is,
+
P[p>a] = 1 - p[p~~] = 1 - F (*) = ; (2.20)
D
v
o
where F+ (a) is the cumulative probability distribution function of the
peak values. The corresponding pd. f. ~(a) is obtained as:

dFP(a) du+
fp(a) = da = 1 (2.21)
~ d:
o
As wilI be discussed later, ocean waves can be considered as a
stationary narrow band Gaussian process with zero mean. It can be
shown that for such a process % and & are statistically independent,
a
i.e., the slope ~ is independent of the magnitude K . Therefore, the
j.p.d.f. is given by:

where the individual variances Jp xL and ~~ are given in terms of the

A
wave spectral density4 S( u )by
w
= S(m) d~ (2.23)
I
o

and
m
2
0.
x
=
fo (J2S ((II) da
(2.24)

4. S (m) represents the one-sided energy spectral density. The +

superscript will be dropped to simplify the notation.

30

+0
Thus, from equation (2.17) and (2.22) the mean rate of positive
crossing of an amplitude of level a is

1
a.
+ 1 x--r{ +) 2
v = e
a 21 (2.25)
x

Due to symmetry of the Gaussian process about zero v;= >:, therefore
O* a
u = 1

x J+ ( ) 2
a r x
x
(2.26)
and +
Vo= v;= ++--
x

Since a wave spectral density is a relatively narrow-banded

spectrum, its apparent (mean) circular frequency W< is
1
m T
u 2 S(u) d~
/
u =2N& i = o
e m (2.27)
ax
S(OJ) dm
1o
[!
Furthermore, the p.d.f. of the peaks from equations (2.21) and (2.26)
is given by

-+(~) 2
fp(a) = ~ e x
(2.28)
x

which is the Rayleigh distribution with a parameter .%

Both equation (2.25) and (2.28) are important results for ocean
waves. Equation (2.25) gives the average number per unit time
(mean rate) of crossing a wave amplitude of level a and equation
(2.28) gives the p.d.f. of the peaks. In general, the following
important result was obtained: The peaks of a stationary,

31

+/
 narrowband Gaussian process (e.g. ocean waves) follow a Rayleigh
distribution with parameter EX given by
m
2
= E [x2] = S(a) do
x = ax I o
= area under the energy (mean square) spectral density (2.29)

2.1.6 Typical Wave Data

From sea data, oceanographers found that:

1) Over a short period of time (one hour) the wave records can be
assumed to be stationary, relatively narrow-band random process.

2) At any time t the elevation of the wave surface from the

mean follows a normal (i.e., Gaussian) distribution given by (see
Figure 2.13).

(2.30)

where rx = standard deviation

a%
2 = /+= X2 f(x)dx = E[x2]
-m

Notice that the variance is equal to the mean square since the mean
of the wave elevation E[x] is taken equal to zero.

3) The peak amplitude is found to follow closely the Rayleigh

distribution given by

2
f~a) = Z . e-~~ a>o (2.31)
E&
where EXis a parameter (see figure 2.14).

32
It has been shown in the previous section that for a narrow-band
normal process, with zero mean, the distribution of the peaks follows
a Rayleigh distribution with parameter

Eg = E [xz] = ~k~= mean square of the process

= area under the energy spectrum

I f(x)

x
Figure 2.13. Probability Density Function of Wave Elevation

This shows the importance of these spectra. Several w a v e

statistics regarding wave amplitudes can be derived from the
Rayleigh distribution. For example:

Average wave amplitude = 1.25 ~

Average wave height = 2.5 ~

Average of 1/3 highest waves (significant wave height) = 4.o~

Average of 1/1 O highest waves = 5.1 ~X (double amplitude)

Average of 1/1000 highest waves = 7.7 ~ ~X (double amplitude)

33
 $(d)

Figure 2.14. Rayleigh Distribution of the Peaks.

2.1.7 Typical Wave Spectra

It is useful for design purposes to obtain many representative

spectra for different wind velocities or significant wave heights. A
number of formulations are presented next.

1. Pierson-Moskowitz (1964)

Moskowitz [2.4] selected spectra from available data in the

North Atlantic and grouped them in a family of five wind speeds
equal to 20, 25, 30, 35, and 40.

Pierson arrived at the following analytical formulation to fit

these spectra (see Figure 2.15).

2
s (w). = +.
#(g/v@4
w (2.32)

where:

s (u) = spectral density (energy spectrum)

d = frequency, rad/sec .

34
 -3
oi. 8.1 x 10

= 0.74
!

1 = acceleration of gravity, ft/see?

V = wind speed, ft/sec

Any other consistent units could be used in (2.32)

s (w)
p

i
I
/ \

I w

.

Figure 2.15. Pierson-Moskowitz Spectrum.

2. Bretschneider Spectrum

The proposed wave spectrum developed by Bretschneider [2.51

. .
can describe developing and decaying seas, unlike Pierson-
Moskowitz spectrum which describes fully developed seas.
Bretschneider spectrum can be written in the form

-5 =-AU-4 (2.33)
s(u) =au

35
where x and are given by
P

5
a=
16 4 ( iL/3 )2
P

P-=+
P

It should be noted that Pierson-Moskowitz spectrum is

completely specified by wind velocity (one parameter) while
Bretschneider spectrum is specified by two parameters; the
significant wave height ~,, and the peak frequency tit.
,,2n

3. The International Ship Structure Congress Spectrum (1967)

The ISSC [2.6] adopted a two parameter spectrum given by

-B*+ (2.34)
S(u) = AB u-s e

where

A= 0.25 ( ~,,> )2
in
B= ( 0.817 ~ )+
T= mean wave period

H? = significant wave height

This spectrum is intended to be used in conjunction with

observed wave heights and periods.

In general, the shape of the wave spectrum depends on:

36

46
 1) Wind speed (most important parameter)

2) Wind duration

3) Fetch (distance over which the wind blows)

4) Location of other storm areas from which swell may travel

It should be noted that waves may attain their fully developed

state for winds up to 32 knots. Beyond that, it is unlikely for waves
to attain their fully developed state. For example, according to
Pierson, a fully developed sea would result if a wind of 52 knots
blew for 80 hours over a fetch of 1800 n. miles. Such conditions are
not common.

2.1.8 Directional Spectra:

So far the so-called point spectrum (1-D) has been discussed.

This is obtained from records taken at a fixed point with no
indication of the direction of wave components, that is, no indication
of how much each of the components of the wave in different
directions contributes to the energy (spectrum). Such a
representation is adequate for long-created seas, but a more
complete representation is given by a 2-D spectrum S (w,p ) where
h= angle between wave components direction and prevailing
wind direction (see Figure 2.16).
w

1,

0.
W

0.

trum

37
 The angular integration of this spectrum will yield the same
one-directional spectrum as would be obtained from a record taken
at a fixed point.

The 2-D spectrum is much more difficult to obtain and

sometimes it is assumed that a directional spectrum can be
approximated by two independent functions.

s (Ld,
p) = s (h)) f(p) (2.35)

where f ( ~ ) is called the spreading function and S( u ) is the one

directional spectrum. The spreading function f( ~ ) can be assumed to
be:

f(u) = (2/lr) COS21J

(2.36)

Thus,

= ~m S (u)du - /+n\2 (2\m) ~~~2@~ = fm S (u)du

o -Tr/2 o
that is, integration of the directional spectrum with respect to w and
? = integration of point spectrum with respect to u.

2.1.9 pe* DiStributim of a General Stationary Gaussian Random Process:

In a few cases of - wave spectra (and vessel response) the

narrow band assumption may not be adequate. Therefore in this
section the, probability distribution of the peaks of a stationary
Gaussian (normal) random process with zero mean that is not
necessarily narrow-banded is presented. The following results were

38

4P
first obtained by Rice [2.1] and then used by Cartwright and
Longuet-Higgins [2.7] and their proof is given in reference [2.1].

Instead of the Rayleigh distribution obtained earlier (equations

(2.28) or (2.31) ), the p.d.f. for the peaks is given by:
2
a

fp(a,4) =
4-=
e
2camo
+ w-+ 0
e : .
.

+ kc
- *
[ 1 (2.37)

and, by integration, the corresponding cumulative distribution

function of the peaks is given by

where
1
t
2= bandwidth parameter = 1 - ~
ml
o m,

9
*(U) =

.0
I
&
-+dz

m
m
=
1 o
n S(w) d- : nth moment of the spectrum n=0,2,4

It should be noted that m. is equal to the mean square or

variance of the process ~k~ .

As & approaches zero the process approaches the ideal narrow

band process and both equations (2.37) and (2.38) reduce to the
p.d.f. and cumulative distribution function (c.d.f.) of the Rayleigh
probability law. On the other hand, as E approaches one the two
equations reduce to the Gaussian (Normal) probability law, that is,
the peak distribution reduces to the distribution of the surface itself.
It should be noted that both positive maxima and negative maxima
are included in equations (2.37) and (2.38) as can be seen from
figure 2.17.

39
 positive fp(a,i)
maxima negative
maxima

i=l
(Gaussian
p.d.f.)

Figure 2.17. P.D.F. of the Peaks of a General Gaussian Process

2.2 Dvnamic Loads and Response of a Floating Vessel Considered as a

Rigid 130dy:

The objective now is to determine a floating vessel response

(output) for a given state of ocean waves (input) probabilistically
described as discussed in the previous sections. In order to do this,
some preliminary definitions are necessary. .

A fixed parameter or time invariant system means that if a

deterministic input x(t) produces an output y(t) then x(t+ ~ ) produces
y(t+ Y ) where Y is a time shift. A linear system means that if xi(t)
produces yi(t), then x(t) = al xl (t) + alx~(t) produces y(t) = at y, (t) +
az YZ(0 where a I and a.2 are constants. Such a system is governed
by a set of linear differential equations with constant coefficients.

Some of the properties of such a linear system include:

If the input x(t) = ei~t then y(t) = Aei~t

where A does not depend on time t. If the input has an

amplitude X(O) dependent on the frequency o, then the output
amplitude Y(u) will also depend on co, i.e.

If x(t)= X(u) ei~t

then y(t) = Y(u) ei~t

40

5(9
 (2.39)

H( CII) is called the transfer function or frequency response

function or Response Amplitude Operator (RAO). The last
terminology usually refers to the modulus IH(co)1. The transfer
function is thus an output measure of unit input amplitude.

?. z.1 Random Input- Output Relations for Floatin~ Vessels:

OCEAN wANES YE5SEL +mrm SHIP RESPONSE

x(t) Y(t)
ASSUMED
;~(u) (INPIJT ) LINEAR (ouTPUT )
x
s;(u)

Figure 2.18. Input/output System.

We will proceed now to determine the ship response (output)

for a given state of ocean waves (input). Since the input X(t) is
random, we ekpect to. get a random output Y(t) (see Figure 2.18).
Some statistics of the random output may then be useful for design.
A floating vessel , response could be vessel motions such as pitch,
heave, roll, etc., the corresponding velocities and accelerations, .
bending moments (vertical and horizontal), torque or shear forces.

In order to determine the vessel response, the following

assumption is made (introduced first by St. Dennis and Pierson [2.8]).
The ship is assumed to behave linearly so that the response can be
described by the superposition of the response to all regular wave
components that make up the irregular sea.

It should be noted that in very severe seas certain responses

may not be linear and non-linear analysis must be conducted.

Using the linearity assumption, the following conclusions will

be stated and the proof can be found in [2.2]:

41

5/
 1. If the excitation (wave input) is a stationary random
process, the response (output) is also a stationary random
process.

2. If the input is a normally distributed random process, the

output is also a normally distributed random process.

3. If the mean of the input process is zero, the mean of the

output process is also zero.

4. If the input is an ergodic process, the output is also an

ergodic process.

Notice that if the input process is nmow band, the output is

not necessarily a narrow band process. For ocean waves, we have
assumed over a short period of time a stationary normal random
process with zero mean. The process could be completely
characterized by the spectral density Sx (co). The area under the
spectrum is related to the. mean square of the process, therefore
certain averages such as average wave height, average of 1/3
highest waves, etc., can be determined. (The subscript x in the wave
spectrum Sx (m ) is used in order to distinguish it from the output
(response) spectrum S~(o) ).

Using 1, 2 and 3 above, it can be concluded that a floating

vessel response is a normally distributed, stationary random process
with zero mean over a short period of time. Again, just as in the
input waves, if the spectral density of the vessel response is
obtained, the mean square value, certain averages and other
statistics of the vessel response can be determined.

It is thus important now to find the relationship between the

wave spectrum and the respons~ spectrum. For linear systems, it
can be shown [2.2] that this relation is generally given in the form

Sy (w) = Ill (w) 12SX(U) (2.40)

where H ( w ) is the frequency response function or the transfer

function afid its modulus IH ( W )1 is the Response Amplitude Operator
(RAO) - see also equation (2.39).

42

5.2
 Equation (2.40) gives the input-output-relation in a frequency
domain, i.e., between the spectra of the input (waves) and the output
(vessel response). Similar relation can be obtained in the time
domain between the response time history y(t) and the wave time
history x(t). This relation as well as other relations in the time and
frequency domains are developed in [2.2]. The important results are
given here as follows.

The response of a vessel y(t) (time domain) for any arbitrary -

known wave excitation x(t) is
w

y(t) = (2.41)
J x(t - 0) h(9) d9
D
and the mean of the response E[y(t)] in terms of the mean of the
stationary excitation E[x(t)] (if different from zero) is

E[y(t)] = E[x(t)] ] h(6) d9 (2.42)

-o
where h(0) is called the impulse response function which is the
response of the vessel due to unit excitation. Notice that E[y(t)] is
actually independent of time since E[x(t)] is independent of time if
the process is stationary.

The relation (time domain) between the autocorrelation

functions of the response Ry( 2 ) and the wave Rx( ? ) is given W
[2.2]. - +m

Ry(lY)= Rx (~+0@2)h(el)h(02 )dE)l de2 (2.43)

1[
.@ -m
The impulse response function h(t) and the ~transfer function
H( d ) are not independent. In fact, together they form a Fourier
pair. +*
LA
h(t) = ~ H(ti)e dw (2.44)
1--
and . k
b-
H(ti) = %(t) e dt (2.45)
J
Load and re~~onse determination for floating vessels is usually
done in frequency domain. Therefore emphasis will be given on this
in the following sections.

The frequency response function H( @ ) or the RAOS are

functions that give the vessel response to a regular wave of unit

43
amplitude. For example, if the response under consideration is the
bending moment, then the bending moment can be calculated for the
vessel in regular waves of different frequencies and for different
vessel speeds. The RAO curves would appear as shown in Figure
2.19.

Af 01 FFE12ENT
5H IP sPIEEIZ:)

Figure 2.19. Response Amplitude Operator

Notice that the ordinate of IH (m)12 is the square of the bending

moment per unit wave amplitude . This can be given in the
nondimensional form 3

where
f= water mass density
g = acceleration of gravity
B= vessel beam
L= vessel length

The RAO of other responses such as shear force could be

obtained in the form lH(m12 = (V/~)2 or the nondimensional form

IH(w)IZ = (~}2

44
where V is the shear force.

In general the RAOS can be obtained either from:

a) Calculations using the equations of motion of the ship

b) Towing tank experiments

Each of these will be discussed briefly in the following.

The general dynamic equations of motion of a vessel in regular

waves can be obtained by applying Newtons Law of Motion for a
rigid body. If the origin is taken at the center of gravity of the body,
then

and
iii
where
= velocity vector
= force vector
m= mass
n= moment acting on the body
G= angular velocity vector
I = moment of inertia about the coordinates axes

The first of these equations give the three force equations in

the x, y, and z directions (surge, sway, and heave equations). The
second gives the three moment equations about the x, y, and z axes
(roll, pitch and yaw equations).

These general six coupled differential equations for the six

degrees of freedom are highly nonlinear and difficult to solve
exactly. However, approximate solutions after decoupling some of
the motions from each other and going through a linearization
procedure are available, for example, in [2.9] and [2.10]. The
decoupling of the equations is usually done by decoupling the
motions in a vertical plane (surge, heave, and pitch) from the rest
and neglecting the surge degree of freedom. The solution of these
equations permits the calculation of the vessel motions and
accelerations in regular waves of different frequencies. For further
information on this subject, see references [2.11, 2.12, 2.13, 2.14,

45

5s
2.15]. Once the vessel motions and accelerations are determined, the
shear force and bending moment (or any other loads) can be
computed. The values of these responses (including Ioads) due to
waves of unit amplitude and different frequencies give the required
RAOS. There are several computer programs to perform these rather
lengthy computations, for example, [2.16, 2.17, 2.18, 2.19].

The RAOS can also be determined by simply running a model

in regular waves at various speeds, headings and wave frequencies.
The model has to be scaled properly to represent the ship mass
distribution and structural geometiy. The model motion, velocities,
accelerations, shear forces, bending moment, etc., are then measured
and plotted versus the wave frequency [2.20]. If the bending
moment needs to be measured at the midship section only, then one
may use a rigid wooden model jointed at the midship section with a
dynamometer calibrated to read the bending moment acting on a bar
connecting the two parts. If the bending moment is desired at more
than one location, then a segmented model is usually used with a bar
equipped with several strain gauges [2.21].

With the RAO determined, equation (2.40) can be applied

to determine the energy spectrum of the response in long-crested
seas. This is usually represented graphically, as shown in Figure
2.20, for the bending moment case taken as an example.

sJd

Figure 2.20. Response Spectrum.

Equation (2.40) indicates that ordinates of the bending moment

spectrum are obtained by multiplying the ordinates of the wave
spectrum by the square of the ordinate of the response amplitude
operator. Since over a short period of time the response is a

46
 stationary normal process, then the response spectrum characterizes
the process completely. If the resulting wave bending moment
spectrum is narrow-band as is the case usually, then the amplitudes
of the wave bending moment follow the Rayleigh distribution
(equation 2.28) with a parameter Ey related to the area under the
bending moment spectrum.

Ey = area under the energy spectrum of the bending

= J ;:)%,;.
o
Some statistics of the bending moment can be thus obtained:
average amplitude of bending moment = 1.25 ~y

average of 1/3 highest amplitude of bending moment = 2.0 {Ey

average of 1/1O highest amplitude of bending moment = 2.55 / Ey

In general if the response spectrum is not narrow-band, then

the peaks (including negative maxima) will follow Rice distribution
given by equations (2.37) and (2.38) with the band width parameter
~ determined from the moments of the response spectrum. It
s hou1d he noted that the assumption of a narrow-band spectrum
produces more conservative results and simplifies the analysis
considerably.
.
2.2.2 Frequency Mappin~ .-

The response spectrum discussed, Sy ( u ), is not the spectrum

that would be obtained by records taken of bending moment aboard
a vessel and analyzing them. This is because, when the vessel is
moving, the waves are encountered at different frequencies to their
absolute frequencies. Consider a vessel moving with velocity V and
heading 0 in regular waves of frequency w (see Figure 2.21).

47
 Figure 2.21. Heading Angle in Regulti Waves.

The wave velocity C in the x-direction = ctI/k, where k = wave

number = 211/ A . The relative velocity between waves and the ship =
C-vcose= encountered wave velocity. Therefore, the encountered
wave frequency Ole= (C- Vcose)k or we= co- kV cos 9. For gravity
waves

where ~ is the wave length

then, me = O-J!L2 Vcose (2.46)
g

Now the wave spectrum can be plotted to a base of the encountered

frequency rather than the wave frequency. However, a change in
the base will necessitate a change in the ordinate of the spectrum
such that [2. 10]:

Sy (d)b= Sy (+) $We (2.47)

Therefore, on integration

co m
Ey =
J
Q
Sy (UJ)&=
JSy
c
(@e)d@e

48
 But from equation (2.46) $@e= [1 -@V cos e] $U
substituting in equation (2.47) a

i
s@) = Sj(ti) (2.48)
[i. T
~ v Lzuej

Thus, for a given Sy ( @ ) and vessel velocity V, SY (we) Cm be

obtained from equation (2.48).

The input-output relation in the frequency of encounter

domain becomes:
Sy (WJ = y(u@l& Sx (~

where ~ (me) is the response amplitude operator obtained as a

function of the frequency of encounter, and the response spectrum
S y ( LJe) relates to records taken on board the vessel.
m .m
It should be noticed that Ey = J
o
Sy (u) du= I
0
Sy (~e) dti~

2.2.3 ~

In short-crested seas, when two-dimensional wave spectrum

S( @ , p ) is used, the input-output relation becomes:

Sy (utp) = Iy(ti, d - ~ )12 Sx (~ ,p ) (2.49)

where p = angle between wave component under

consideration and the prevailing wind
direction.
d-~ = angle between the vessel velocity vector and
wave component (see Figure 2.22).

The response of the vessel for all wave components can be then
obtained by integration over p

49
 VESSEL VELOLITY VECTOR
/
WAVE COMPONENT
//

PREYAILIN6 WNO
4
DIRECTION

Figure 2.22. Angle Between Wave Component and Prevailing Wind

Direction.

It should be noted that several computer programs are

available to determine the response spectrum of a vessel based on
the above or similar analysis. Such computer programs include
,~ [2.1 61, SpRmGSEA [2.17], ~~ ~ Sfip W+qm p~ [2.181.

2.3 Long-Term Prediction of Wave Loads:

In the previous discussion, one of the major restrictions has

been the assumption of stationarity which limits the validity of the
analysis to short periods of time. This leads to Rayleigh distribution
of the peaks for narrow band spectra. For design purposes,
however, it may be important to determine the distribution of the
wave load peaks over long periods of time (years) in order to
determine design values of the load.

The long term distribution can be determined by adding

statistically several short-term (Rayleigh) distributions, or, by taking
records of wave loads and determining what probability distribution
gives the best fit of the data. Several statistical methods can be used
to estimate the parameters of the candidate distributions and tests
are available to examine the goodness-of-fit and to determine which
distribution fits the data best.

Several investigators, [2.22, 2.23, 2.24, 2.25, 2.26] examined

long-term wave loads data with the aim of determining the long-
term distribution of the peaks. It was found that the Weibull
distribution is general enough and fits well the wave bending
moment data on ships. The p.d.f. and c.d.f. of the Weibull
distribution are given by: #
+,ttf

. (+f-J e
(x)= (++ x~o (2.50)

50
and - (%$
F(x) =l-e X>o (2.51)

where ~ and k are parameters to be determined from the wave load

data (e.g. strain gages installed on deck of a ship).

It should be noted that the Weibull distribution is a generalized

Rayleigh distribution and if one inserts ~ = 2 and k =~ in equations
(2.50) and (2.5 1) one obtains directly the Rayleigh distribution p.d.f.
and c.d.f. (see equation (2.31) ).

It has also been shown that in many cases of long-term wave

load data, the parameter ~ of the Weibull distribution is
approximately equal to one. When ~ = 1, the Weibull distribution
reduces to the Exponential distribution given by:
+,~)
f(x) =*e (2.52)
- (WA)
F(x)= l-e (2.53)

where ~ = k = mean or expected value of the wave load amplitude.

Section All of ~ 1 desaibes hcw to detemine the ~alEterS k and 1
and h ti construck and use a probabili~ ~.
2.4 Prediction of ~ Wave Loads:
If the wave loads acting on a marine structure can be
represented as a stationary Gaussian process (short period), then at
least four methods are available to predict the distribution of the
maximum load. These methods are, developed for application to
marine structures and are given in more detail in [2.27]. In the first
method the peaks are assumed to be statistically independent and
identically distributed, and the extreme value distribution of the
largest in N-peaks is determined using classical order statistics. In
the second, a discrete point process is assumed in order to determine
the asymptotic type-I distribution based on Rices [2.1] initial
distribution. Cramers procedure [2.28, 2.29] can be used for
determining the resulting aysmptotic distribution. Conventional
uncrossing analysis may be used in the third method for determining
the extreme value distribution. Finally a two-stage description of the

51
random process which leads to an extreme distribution derived by
Vanmarcke [2.30] is the basis for the fourth method.

Each one of these methods will be described briefly in the

following sections.

A. Distribution of the largest peak in a sequence of N neaks using

order statistics

The distribution of the largest peak in a sequence of N peaks

can be determined using standard order statistics. Consider a
sequence of random variables 21,22, . . . Zn representing the peaks of
a load on a marine structure. Assuming that these peaks are
identically distributed and statistically independent, the cumulative
distribution function (calf) of the largest one using order statistics is
given by [2.31]:

F (Z) = P rnax (z*, Z2, ... . ~, s z] = [FZ(%.)]N (2.54,

$ [
<,
where FZ (z, G) is the initial cumulative distribution function of the
load peaks (maxima) and E is the spectral width parameter defined
as:

E = 1 - m22
o 4

+W

rnn =
J con S(u) do) , n = 0,2,4 (2.55)

The probability density function (pdf) of the largest peak is

determined by differentiating equation (2.54) with respect to z, thus:

%(z)
= [FZ(Z4N-* zzE) (2.56)

where fz (z,&) is the initial p.d.f. of the load peaks. For any band-
width load process, Rices distribution should be used as the initial

52
distribution, thus equations (2.37) and (2.38) hold for fz (z,&) and FZ
(z,E), respectively.

Based on the above analysis, the expected value of the

maximum load peak in a sequence of N-peaks was determined by
Cartwright and Longuet-Higgins and is approximated by:

E[max (zl,z2, ,.. zn)]~

0%
EC

where C = 0.5772 = Eulers constant .

The extreme load peak with a probability of exceedence a is given by

[2.32]:

z=
a [(2moln N + h
[
l/ln(~
) 1}1+. (2.58)
which is independent of 5 (for mall a ).

(B) Asymptotic type I distribution

It is known that as the number of peaks N increases without

bound a limiting or asymptotic form of the extreme value
distribution (equations (2.54) and (2.56)) is reached. The asymptotic
form of an extreme value distribution does not depend, in general, on
the exact form of the initial distribution; it depends only on the tail
behavior of the initial distribution. The parameters of the
asymptotic distribution depend however on the exact form of the
initial distribution.

In reference [2.27], Cramers method was used to derive the

asymptotic distribution based on Rices distribution as an initial
distribution. The derived extreme value c.d.f. is

53

n
 F (z, G) = exp
%

that is, the asymptotic form is double exponential and the cumulative
distribution itself depends on N. ms is the mean value of the load
if different from zero.

Several years after the appearance of Cramers book Gumbel

[2.3 1] classified the asymptotic distribution of extremes in three
types: (type I) a double exponential form, (type II) an exponential
form, and (type III) an exponential form with an upper bound.
Convergence of an initial distribution to one of the three types
depends largely on the tail behavior of the initial distribution. An
initial distribution with an exponentially decaying tail in direction of
the extreme will converge to type I asymptotic distribution, i.e., the
double exponential form.

Gumbels analysis and classification provide another method

for deriving the asymptotic distribution and may be in a form easier
to handle than that given by equation (2.59). The cdf of type I
asymptotic form as given by Gumbel is:

F
%
(Z) =exp-e [
- aN{z I.@
1 (2.60)

where u ~ is the characteristic largest value of the initial variate Z

and d ~ is an inverse measure of dispersion of Z ~ . These parameters,
UN and d ~ , have to be determined and depend on the form of the
initial distribution.

The corresponding pdf is given by:

- aN (z-UN) -%(2-%)
f (zJ=~Ne .exp -e
+ [ 1 (2.61)

54
The mean and standard deviation of the extreme value Z ~ are given,
respectively, by:

(2.62)

K
Zn = J-
6 aN (2.63)

The parameters u ~ and d ~ were determined in [2.27] for Rices

distribution given by equations (2.37) and (2.38) as an initial
distribution.

The results for u ~ and d~ are:

N
=m+
s
(2.64)

(2.65)

where
N - S
a=.
Em
r o
and
(2.66)

55
The plus sign in equation (2.64) should be used if the mean value ~
is positive in order to obtain the larger characteristic value. It
should be noted that both A and ~ contain u ~ as defined in (2.66);
therefore, an iterative procedure must be used for determining u ~ .
To start the iterative procedure an initial value for u ~ is necessary
and may be taken as u ~= rn~t~~. The corresponding values of @ , ~ ,
+(-u) and @ ( ~ ) can then be determined. Equation (2.64) is
then checked to see if the right side is equal to the left side,
otherwise a new value of u ~ equals the right side of equation (2.64)
should be used in the second step of the iterative procedure. Three
or four steps are usually sufficient for convergence.

(C) Extreme value distribution based on uncrossing analysis

The distribution of the largest peak can be determined from

uncrossing analysis of a time history of a stationary random process
instead of the peak analysis presented above. For example, the
number of N peaks can be changed to a time interval T in the
uncrossing analysis and the problem of determining the
characteristics of the largest peak in N peaks becomes that of
evaluating the characteristics of the maximum crest of a stationary
ergodic Gaussian random process X(t) during a period T. The
assumption of the statistical independence of the peaks is usually
replaced by the assumption that uncrossing of a level x by X(t) are
statistically independent. This leads to the Poissons uncrossing
process which is true only in the asymptotic sense (as x - e ; T_= ).

From uncrossing analysis it can be shown (see for example

[2.1]) that the probability that the largest value is less thar. 5. mrtain

level x during a period T is given by:

P max(X(t);
[
OSt S~)sx
1s (2.67)

where

56
 2

--( )
lx
- m~
2%
r (2.68)
<=voe

and

v = ~~ ltsec
O ~ J m. (2.69)

Therefore the cdf of the largest X is

\~wlj
FX(X) = exp
I -vOTe

(2.70)

that is, it has a double exponential form although quite different

from equation (2.60) with u ~ and ~fi given by (2.64) and (2.65).

(D) Extreme value distribution based on a two-state description of

a random process
.
.
Vanmarcke [2.30] estimated the probability distribution of the
time to first passage across a specified barrier for a Gaussian
stationary random process. In his analysis he considered a two-state - .
description of the time history X(t) relative to the specified barrier.
Based on his results the distribution of the extreme value may be
determined from:

57
 Fx (x) = exp -v. T -e
(2.71)
L

where q is a band width parameter defined as

(2.72)
()~q~l

Distributions given in (A), (B), (C) and (I)) above, are valid for a
load process represented by a stationary Gaussian process of any
band width. Reference [2.33] shows the corresponding equations for
the special cases of a narrow-band process ( E = o or q = o) and a
wide-band process ( & = 1 or q = 1). It should be noted that the
narrow-band case gives a conservative estimate of the extreme wave
load distribution and the resulting equations may be used for values
of ~ Up tO 0.60 shce they are insensitiveto E in the range O to 0.60.

Application Example and Comparisons:

The extreme value distributions of the wave loads discussed

above differ from each other in their basic derivation and underlying
assumptions. The forms of their equations are drastically different
as can be seen by comparing equations (2.54), (2.59), (2.60), (2.70)
and (2.71). It would be interesting now to compare some typical
results obtained from the different methods when applied to a
marine structure. For this. purpose a tanker of length = 763 feet,
breadth = 125 feet and depth = 54.5 feet is considered. We will
compare the distribution of the extreme wave bending moment
acting on the tanker under a storm condition specified by a
significant wave height of ~~.75feet and an average wave period of
11.5 seconds. The storm is assumed to be stationary under these

58

b!?
conditions for a period of one hour. The following parameters were
computed for an earlier application given in [2.33]:

Still water bending moment (full load)

ms = 669,037 ft-tons

RMS of wave bending moment

m. = 286,300 ft-tons
r

Average wave moment period = 13. o seconds

Band width parameter of wave moment spectral

density ~ = 0.364

Number of wave moment peaks in one hour

N=, 60 X 60
= 276.9
13a

The application given in reference [2.33] shows that if & is

assumed to be zero (ideal narrow-band) instead of the 0.364 given
above, the resulting error in the expected maximum wave bending
moment in N peaks s less than 0.5 percent. This gives an indication
that for E = 0.364, it is sufficiently accurate to use the ideal narrow-
band equations for our comparison.

Using this assumption and the above values forms ,mO

F and N,
a comparison is made of the distribution functions of the four
extreme value distributions as given in the preceding sections (A),
(B), (C) and (D).

The results of the comparison are shown and plotted on a

standard extremal probability paper and on a regular graph paper in
figures 2.23, 2.24 and 2.25.

59
 Based on these results, one surprising conclusion can be drawn.
All extreme value distributions of the wave loads considered produce
similar results even though their basic assumptions and derivations
differ drastically. In fact, if one inspects the equations representing
the cumulative distribution functions of these distributions one sees
that these equations are not similar in form and may conclude
erroneously that they would produce very different results.

The extreme distributions based on the largest peak in N peaks

(distribution A), uncrossing analysis (distribution C ) and a two-state
description (distribution D ) produce almost identical results as far as
the probability of exceedence is concerned as can be seen by
inspecting figures 2.23 and 2.24. The asymptotic type I distribution
(distribution B) results in slightly higher values of probability of
exceedence. This is to be expected since the asymptotic distribution
is an upper bound extreme distribution and becomes more accurate
as the number of load peaks approaches infinity. In the example
shown for the tanker, the number of wave bending moment peaks N
is approximately 277.

As an example of the differences between the asymptotic

distribution and the other distributions, the probability of
exceedence of a total bending moment of 2,069,000 ft-ton (including
still water bending moment of 669,000 ft-ton) is 0.006 according to
the asymptotic distribution (B) and 0.002 according to the other
three distributions (A, C, and D).

60
 .001.01 .10.20 .30 ,50 ,70.m .90 .95.97.98 S .595 .997s 999
Ptiabilily CDF

Figure 2.23. Standard Extr~mal Variate - Bending Moment on a

Tanker.

co$
10 lhuONM, A C, ti O

00

09

47

0.6

0.$

04

02

02

0.1

0.0
--,- .-*,- -ti---
--m

Figure 2.24 - Cumulative Distribution Functions of Extreme Bending

Moment on a

61
 pd. f.

xlo~
I DMnbuhonsA andC
5,0
............... DMribuuonB

4.0
1

3.0

20

10

0.0 I
700 900 llIJO 1300 1~ 1700 1goo 2100 2300 2500 27OO I1OOO

Bending Momcm (It-tons)

Figure 2.25. Probability Density Function of Extreme Total Bending

Moment on a Tanker

2.5 ~ ilities:

The probability that an extreme value of an event (say wave

height x) will not be encountered during the life L of a marine
structure is called non-encounter probability NE(x). This, in
general, is given by [2.34]:

NE(x) = P [ no exceedence of x occurs during life L]

= P[Xm~~X] = [Fx(x)]L
(2.73)

where Xmu = maximum value during life L

L = life in years
FX (X) = distribution function of annual maximum.

The waiting or return. period, R, is the average length of time

between exceedence. Thus one may speak of a 100 year wave height
or 50 year wind velocity.

The waiting period in years has a probability law given by

62

72
 P[w=w] = FX w-l(x) [1- FX (x) ]

and therefore, the average waiting period, i.e., the return

period R is

R = E[W] = [ 1 -Fx (X) ]-1 (2.74)

The relation between the non-encounter probability NE(x)

and the return period R can be determined by eliminating FX (x)
from equations (2.73) and (2.74) , thus,
-I L
NE(x) = PIXmu Sx]= [l-R] (2.75)

If R = L, then NE(x) ~ e- 1

The probability of exceedence in this case = 1 - e-l = 0.632,

that is, there is a high probability (0.632) of exceeding the event
with a return period L during the L life years of the ~tructure.

In selecting return periods, one must distinguish between an

annual interruption of operation of the structure (L = one year) and
ultimate failur~ during life time ( L = 20 to 30 years). In the former
case a return period. R = 10 years may be adequate. Using equation
(2.75) with L = 1 and R = 10 we obtain a nonencounter probability of
90%. If R is increased to 100, the non-enco~nter probability becomes
99%.

In the latter case where failure during life of, say ,20 years, is
considered, and the return period is 100 years, then the non-
encounter probability, from (2.75) is 81.8%. If the return period is
increased to 1000 years; the non-encounter probability becomes
98.0%.

63
2.6 Stochastic co mbination of Loads on a Marine Structure

Undoubtedly, there are certain similarities between

decomposing ship response records of full-scale measurements into
their basic components and combining analytically calculated
components to obtain the total response. Since decomposing full
scale measurements can be done with a certain degree of success, it
is possible to invert the procedure in order to compute the combined
response from the analytically determined components.

In this section, a brief discussion is given of the decomposition

of full-scale records into their basic components. In the following
section, a method is presented to combine analytically-determined
response components.

A. Decompositions of measured records into their basic

components

A typical measured stress time history of a bulk carrier is

shown in figure 2.26 (from reference [2.35] ). Usually, such a record
consists of a rapidly varying time history of random amplitude and
frequency, oscillating about a mean value. The mean value itself is a
weakly time-dependent function and may shift from positive to
negative (sagging to hogging). The two dominant factors which affect
the mean value are:

1 The stillwater loads which can be accurately determined

from the loading condition of the ship floating in
stillwater.

2 The thermal loads which arise due to variations in

ambient temperatures and differences in water and air
temperatures.

A closer look at the rapidly varying part shows that it also can
be decomposed into components. Figures 2.27 and 2.28 illustrate

64

74
records taken over shorter periods of time (larger scale). TWO main
central frequencies appear in these records. The smaller central
frequency is associated with the loads resulting from the motion of
the ship as a rigid body (primarily heave and pitch motions). This
lower central frequency is, therefore, close in magnitude to the wave
encounter frequencies for wave length nearly equal to ship length.
.* *

.>

Fig. 2.26 Typical voyage variation of midship vertical bending stress

for a bulk carrier. From reference [2.35]

Figure 2.27. Decomposition of a stress time history of a Great Lakes

vessel into low and high frequency components.

65
Figure 2.28. Decomposition of a stress time history of an ocean going
bulk carrier.

The higher central frequency is associated with loads resulting

from the two-node mode response of the ship when it vibrates as a
flexible body. This higher central frequency is thus close to the two-
node mode natural frequency of the ship. The high frequency
response itself can be due to springing of the flexible ship when
excited by the energy present in the high-frequency wave
components as shown in figure 2.27. It can be due, also to the
impact of the ship bow on the water as the ship moves into the
waves, i.e., slamming (possibly together with low-speed machinery-
induced vibrations), see figure 2.28. Though springing and
slamming may occur simultaneously, it is unusual to see records
which exhibit both clearly. These two responses can be distinguished
from each other by inspecting the records envelope. In general, a
decaying envelope (see figure 2.28) indicates a slamming response
whereas a continuous envelope of varying amplitude, as shown in
figure 2.27, indicates a springing response.

The rigid body and the high frequency responses do not always
occur simultaneously in the same record. Quite often only the rigid

66
body response appears in a record; particularly, in that of a smaller
ship which has a high two-node mode frequency. Occasionally, only
the high frequency springing response appears in a record when a
ship is moving or resting in relatively calm water. In particular, long
flexible ships with low natural frequencies as those operating in the
Great Lakes do occasionally exhibit such records when operating in
calm water or in a low sea state composed mainly of short waves.
Under these conditions, a long ship will not respond as a rigid body
to the short waves, but the two-node mode frequency of the hull can
be sufficiently low to be. excited by the energy content of these short
waves. Figure 2.29 (from reference [2.36]) shows a measured
response spectrum of a large Great Lakes vessel where the response
is purely in the two-node mode S and higher frequencies with no
rigid body response appearing in the spectrum. The figure shows
that response at higher modes than the two-node mode can be
measured, although small and relatively unimportant in most cases.

Slamming response on the other hand never occurs separately

without rigid body response since, obviously, it is a result of the rigid
motion of the ship in the waves.

13 ?
:
?.
z

\
10
~mMs. .*0*sl

5$
I* Mom

2-WOC ,

G~
o 0.3120 0,7215 09165
FREOUENCY!HERTZI

Figure 2.29 Stress response spectrum of a large

Great Lakes vessel. [2,36].

5 The two-node mode is labeled in the figure as the first mode.

67
 Two main steps should be used in the procedure for combining
primary responses of a vessel.

Step 1 To combine the low frequency wave-induced

responses (rigid body) with the high-frequency
responses (springing or slamming).

Step 2 To add the mean value to the response resulting from

Step 1. The mean value consists of the stillwater and
the thermal responses.

SwLL

Consider an input-output system in which the input is common

to several components of the system; it is required to determine the
sum of the individual outputs. Here, the input represents the waves
which can be in the form of a time history if a time-do-main analysis
is sought, or in the form of a sea spectrum if a frequency-domain
evaluation is preferred.

r
--+_ ____ ____ ____ _l

L_-___ ---- . . .. ~i,; - i

Figure 2.30 Schematic representation of a multiple system with

common input.

68

-78
 The components of the system represent the components of the
load response of the ship to waves, e.g., the low-frequency wave-
induced responses which consist of vertical, horizontal and torsional
moments, the high frequency responses such as the springing loads .
It is required now to determine the sum of these component
responses, i.e., to determine the output taking into consideration the
proper relations or the appropriate correlation of the response
components,

Schematically, the procedure is represented by figure 2.30. In

this figure, n pmallel linem components are considered which have
common input ? (t) and are summed up at the output to form ~ (t).
The output of each system is multiplied by a constant ai (i = 1, 2, ....n)
before summing up all the components at a common node to form
f(t). These constants ai give additional flexibility in the application
of the model and can be used to weigh the contribution of each
linear system to the sum.

In a time domain, the output ~ (t) is given by the sum of the

convolution integral of each system.

,~(t] ~i~lai[~rnhi(?)~(t - T)dT]

o
, * =~~hc(~)%t-T)d?
o
(2.76) ~ -

where
hc(~) = ~ aihi (T)
i=l
(2.77)

hi (z) is the impulse response function of each linear system, i.e., the
response of each linear system to unit excitation multiplied by time
(response to the Dirac delta function). hc (%) is a composite impulse
response function which sums the responses of the individual
components.

69
 It should be noted that the impulse response functions of the
individual components hi (~) may or may not be easy to obtain
depending on the complexity of the system. With suitable
instrumentation, it is sometimes possible to obtain a good
approximation to hi (t) experimentally. For the ship system, hi (~)
can be determined for most load components.

In a frequency domain analysis, a similar procedure can be

used. In fact, since the system function Hi (co) is simply the Fourier
transform of hi (t), i.e.,

Hi(u) = fhi(t)e-j~tdt
o (2.78)

therefore, we can define a composite system function Hc (co) as

H=(u) = ]rnhc(t)e-j(t
cit
o
=i~laiHi (u)

(2.79)

It should be noted that for a single system, the relation

between the input spectrum and the output spectrum is given by the
usual relation:

[Syyfw)li = sxx(~)H*i(~)Hi(~l

= SXX(W) lHi(w) 12

(2.80)

where SXX(O) is the sea spectrum which represents a common input,

[Syy (0) ]i is a response spectrum of an individual load component ,
H* i (co) is the complex conjugate of the system function of a response
component.

The modulus of the individual system function I Hi (U) I is the

response amplitude operator of the individual response components,
i..e.,

70
 [R. A.O.li = [P(u)] = IHi(u)j
i

and, therefore, equation (2.80) represents the familiar relation

between the input and the output spectra of a single linear system.

For our composite system, an equation similar to the equation

(2.80) can be determined for the n-response components and the
weight factors ai as follows

Syy(w) = SXX(W)H*c(@Hc (w)

n
= sxx(~)~ ~ aiaj Hi (U)H*j (w).
(2.81)
i=l j~l

The double sum in equation (2,81) can be expanded such that a final
expression for the total response spectrum Sy y (OI), which combines
the individual response spectra, may be written in the form:

s (u) = [~ ai21Hi(u)[21SXX (u)

YY
i=l

+ [~ f a.a. H. (OJ)H*j(W)]Sxx (u)

i=l j=l 1 J z (2.82)
(i#j )

It should be noted that the first term in equation (2.82)

represents simply the algebraic sum of the individual response
spectra, each modified by the factor ai . The second term, which cm
be either positive or negative, represents a corrective term which
depends on the correlation between the load components as can be
seen from the multiplication of Hi (m) by the complex conjugate of
Hj (o)).

If the system functions Hi (U) do not overlap on a frequency

axis (i.e., disjoint systems), that is, if

Hi(m)H*j (10) = o
(2.83)

71

81
then the second term in equation (2.82) becomes zero and the load
components are uncorrelated. In this case, the total spectrum is
simply the algebraic sum of the individual spectra of the load
components, modified by the factors ai . Furthermore, if the wave
input is considered to be a normal random process with zero mean,
as usually is the case, then the respective output load responses are
jointly normal and are independent. Thus, the total response, in this
special case, is a zero mean normal process with a mean square value
given by:

= (iyy (u) dw
Y 2

i~lai
;lHi
(w)I sXX(w)dw
o (2.84)

In the more general ( and more realistic) case where some or

all of the response components are correlated, the mean square is
given by

= (Syy (u)dw
aY2

(2.85)

It should be noted that in equations (2.84) and (2.85) the mean

square is equal to the variance of the combined response since the
mean value of the wave responses is usually very small. As noted
earlier, the other responses which consist mainly of the stillwater,
and the thermal loads will be added later in the second step of the
analysis m form a mean value for step 1 combined responses.

72
 In connection with equations (2.84) and (2.85) it should be
ei~ that the usual R@ig~ multiplier used to estimate certain
average quantities, such as average of the highest one third,one tenth
response, etc., are not generally applicable in the case of the
combined response, since these multipliers are associated with a
nmrow-band spectrum for which the amplitudes can be represented
by a Rayleigh distribution. It is only in the case where each of the
load component responses is narrow-band and each happen to be
closely concentrated around a common central frequency w;, that it
would be reasonable to conclude the combined response SYY (w ) is,
itself, a narrow-band process.

In the more general case, where the combined response

spectrum is not a narrow-band spectrum, the various statistical
quantities can be determined from a more general distribution (Rice)
which includes the Rayleigh distribution as a special case. The
general distribution is given by equations (2.37) and (2.38).
.

Equation (2.85) can be written in a different form which is

more convenient to use in applications, and which makes it easier to
define the correlation coefficients between the different response
components.

(2.86)

where

Ui= ~lHi(w) ls xx (~)dw m variances

o
or mean squares - of the response spectra of the
individual load components. (2.87)

,,
s~,= correlation coefficients of individual load
d
components defined by,

73

U
 pij =+rnHi(u)H*j (u)SxX(ti)dW
ijo
(2.88)

Equation (2.86) with definitions (2.87) and (2.88) form the

basis for combining the step 1 responses of a ship hull girder in a
frequency domain analysis taking into consideration the correlation
between the response components. If the response components are
uncorrelated, i.e., if p ij = O, the second term in equation (2.86) drops
out and the variance of the combined responses (output) is simply
the algebraic sum of the individual variances modified by the factors
ai. AS discussed earlier, this occurs when the system function of the
various components do not overlap in frequency or overlap in a
frequency range where the individual responses are small. On the
other hand, if the individual components are perfectly correlated, p ij
may approach plus or minus unity and the effect of the second term
of equation (2.86) on the combined load variance a 2 y can be
substantial.

The physical significance of the correlation coefficient can be

further illustrated by considering only two response components for
simplicity. If p 12 is large and positive (i.e., approaching -t-l), the
values of the two response components tend to be both large or both
small at the same time, whereas if p 12 is large and negative (i.e.,
approaching - 1), the value of one response component tends to be
large when the other is small, and vice versa. If p 12 is small or zero,
there is little or no relationship between the two response
components. Intermediate values of p 12 between O and ~1 depend
on how strongly the two responses are related. For example, the
correlation coefficient p 12 of the vertical and horizontal bending
moments acting on a ship is expected to be higher than of the
vertical and springing moments since the overlap of the system
functions in the latter case is smaller than in the former case.

In a time domain analysis, the convolution integral represented

by equation (2.76) with the composite impulse response function as

74
given by equation (2.77) form the basis for determining the
combined load response. The question of whether the time or the
frequency domain analysis should be used depends primarily on
what form the required input data is available. In general, most
wave data and practical analysis are done in a frequency domain,
although in some cases where slamming loads are a dominant factor,
it may be advisable to perform the analysis in time domain.

SEu
In this step, the stillwater and the thermal responses should be
combined to form the mean value for the rigid body motion and
higher frequency responses. The stillwater and thermal responses
are weakly time-dependent variables so that in a given design
extreme load condition they can be considered constants, say over
the duration of a design storm. Therefore, theses two responses can
be treated as static cases and can be combined for one or several
postulated design conditions without difficulty. Alternately, if
statistical data are available for each of these responses, the mean
and variances of the combined response can be easily determined.

The stillwater response can be accurately determined for all

loading conditions tiing computer programs such as the Ship Hull
Characteristics Prqgram. Several postulated, extreme but realistic,
weight distributions can be assumed in the final stages of design, and .
the corresponding stillwater response can be computed.

If a statistical description of the stillwater bending moment is

adopted, data have shown that the general trend assumis a normal
distribution for the conventional types of ships. A sample histogram
based on actual ship operations data for a containership from
reference [2.37] is shown in figure 2.31. A mean value of the
stillwater bending moment can be estimated based on this histogram
for all voyages, or for a specific route such as inbound or outbound
voyages.

75

gr
 m. OF
Oxuulums

*UWO
nvc.
I
t

Lf e 20
!**
Figure 2.31. Histogram of stillwater bending moment of a
containership.

*5I Iol

Figure 2.32. Correlation between measured stillwater bending

moment and air/sea temperature difference.

76
 These mean values together with a set of standard deviations,
which can also be estimated from the histogram, can be utilized to
determine the extreme total moment using a statistical approach.

Since the stillwater response, the rigid body motion response,

and the higher frequency responses are all functions of the ship
weight and its distribution, it is preferred that the combined
response be calculated for a group of selected loading conditions (and
selected temperature profiles).

Primary thermal response is usually induced by differences in

water/air temperatures and by variations in ambient temperatures.
A study of full-scale stress data measured on a larger tanker
indicates that the diurnal stress variations correlate well, as shown in
figure 2.32 with temperature differentials between air and sea.

Taking the North Atlantic route as an example, the average

diurnal change of air temperature is about 10 0F. The total diurnal
change of deck plating temperatures may vary from 1OC to 50 F,
depending upon the cloud cover conditions and the color of the deck
plating. For estimating the thermal loads on a ship hull, the sea
temperature may be assumed as constant. Once the temperature
differential along a ship hull is determined, the thermal stresses can
be calculated, using either a general purpose finite element computer
program or a simplified two-dimensional approach. The maximum
thermal response may be then added to the stillwater response for
certain postulated design conditions to form the mean value for the
low and high frequency dynamic responses.

Although high thermal responses may not happen in high seas,

a heavy swell may possibly occur under a clear sky. Therefore,
several temperature conditions are to be taken into consideration in
determining the combined design response.

77
 Generally, large tankers traveling in oblique seas may
encounter horizontal bending moments of the same order of
magnitude of the vertical bending moments. Therefore, the
combined effect of the vertical and horizontal moments can be
critical under certain conditions. The distribution of the primary
stress in the deck as a result of the combined effect becomes non-
uniform and assumes a maximum value at one edge. The combined
stress at the deck edge, 8C is given- by

(2.89)

where,
.
GC = combined edge stress
.

fiv(t) = the vertical bending moment component

fib(t) = the horizontal bending moment component

Sv,sh = the vertical or the horizontal section modulus,

respectively.

Defining the combined moment as the combined edge stress

multiplied by the vertical section modulus and using equation (2.89),
we can write,

tic(t) = %C(t)sv = hv(t) + kfih(t)

(2.90)

where,
s
Iz=$
h

78
 Applying equations (2.86), (2.87) and (2.88) and extending the
results for the case of a two-dimensional sea spectrum, the mean
square value of the combined response ~zM~ can be written as (see
equation (2.86)):

Mc2=a Mv2 + K20~ ~ Z~v#aHVOm

(2.91)

where G*M~ and ah h are the mean square values of the vertical and
horizontal bending moments, respectively, given by equation (2.87)
as6:

aMv 2= ~ /%JM,u) IHV(W)[2 dwdp

-T o

(2.92)

Mh2
= ;$
~ o
/mSxx(u, U)IHh(w) 1 dwdu

-7 (2.93)

Using equation (2.88) the correlation coefficient is defined as:

= 1
vh
Mvw
77

~ &x(wrU)H, =(w)H*h(m)dwdu
--
2

(2.94)

6 For simplicity of notation, the dependence of the RAOS IHv(co)l and

IHh (m )1 on the ship heading with respect to wave components is
dropped from the notation.

79
 The Response Amplitude Operators IHV(U )1 and IHh (co)l and
more generally, the system functions Hv (OI) and Hh (OI) can be
determined from any typical rigid-body ship motion computer
program.

It should be noted that in equation (2.86), the coefficient al

was taken as unity and a2 was taken equal to

s
=K=$
az h

(see equation (2.90)) in determining equation (2.91). It should be

also noted that the integrand in equation (2.94) contains the
information regarding the phase between the horizontal and vertical
moments and that the integration of such information with respect to
frequency and the angle ~ between a wave component and the
prevailing wave system leads to the determination of the correlation
coefficient as given by equation (2.94). Finally, the root-mean-
square (r.m.s.) of the combined edge stress is given by

.,
(rlns) = ~MC
c
v

(2.95)

where OMc is the r.m.s. of the combined moment given by equation

(2.91).

The r.m.s. of the combined moment and the correlation coefficient

as given by equations (2.91) and (2.94), respectively, were computed
in reference [2.38] for a large tanker of DWT 327,000 tons. The r.m.s.
values of vertical and horizontal moments were computed using a
rigid body. ship motion computer program and combined using
equation (2.9 1) to obtain the r.m.s. of the combined moment. The
results of the calculations are plotted in figures 2.33 and 2.34 versus
the heading mgle. =veral sicjnif icant wave heights
 were considered in order to examine the general behavior of
the responses in low, moderate and high sea states. These results
show that the horizontal bending moments is not small compared
with the vertical bending moment (see figure 2.34). In severe seas,
the maximum response of the combined moment (and the vertical
moment) is in head and following seas. Figure 2.35 shows the
variation of the combined bending moment with the sea state as
obtained by three different methods. The first is based on equation
(2.91) with &MV, @Mh and ~ Vh given by equations (2.92), (2.93)
and (2.94) respectively. The second method is based on equations
(2.91), (2.92) and (2.93), also, but with a comelation coefficient ~ vh
equal to 0.32 obtained from the 1973 ISSC Proceedings (determined
empirically). The third method is based on ~ v h = 0.53 as
determined by averaging the responses in short crested seas as
determined from equation (2.9 1) for all headings and for the three
representative sea states. The mean value of ~ v h obtained in this
manner was 0.53, significantly higher than the ISSC value. However,
the effect of ~ vh on the r.m.s. combined moment is small, as can be
seen from figure 2.35.

As a second application example, the combined vertical and

springing mom-ent will be considered next. Using frequency domain
analysis and using equations (2.86), (2.87) and (2.88) with al = a2 =
1, we obtained the ,mean square value of the combined response
avs2 in long-crested seas as . .

where,

@2 = mean square of vertical bending moment

,*
= Sxx(m) I Hv (u) 12do (2.97)
.J
o
~s2 = mean square of the springing moment

81

.
ql
 m
= Sxx(m) I H~ (o)) 12 (2.98)
f
0
=
Y Vs correlation coefficient

and Hv(m) and H~(u) are the complex system functions of the vetical
wave moment and springing moment respectively.
The system
response functions can be computed using computer programs which
take into consideration the effects of ship flexibility in the response,
such as the springing -seakeeping program SPRINGSEA, [2.3 9].
Applications of equations (2.97), (2.98) and (2.99) to several Great
Lakes Vessels where springing is important is given in references
[2.38] and [2.39]. These equations together with equations (2.76) and
(2.77) for the time-domain analysis have a wide range of
applicability to any two or more dynamic random responses
including, combining primary and secondary~ responses [2.38],
vertical and torsional moments for ships where torsion is important,
high frequency loads with vertical and horizontal moments, etc. In
all of the cases, the coefficient ai must be appropriately determined.
When determining the various statistical averages from the
combined r.m.s. values, the more general distribution given by Rice
for the peaks should be used instead of the usual Rayleigh
distribution.

The above procedure for combining loads is not generally

applicable for combining slamming with wave induced loads.
Reference [2.40] describes a procedure for combining loads on a ship
when slamming is involved as one of the loads.

7 Such as primary inplane loads on grillages due to overall bending

of the hull and secondary lateral pressure arising from the randomly
varying water surface.

82

qz
 20
r

15

10

5-

Head
0 45 90 135 180
Seas
HEADING (DEGREESI

Figure 2.33. Variation of vertical and combined wave bending

moments with heading in long-crested seas.

w o
ul -
Z:
k
:u

15- HIGHSCAS

10 -

Q 5 -
. M&

Head
Seas
o~
o 30 60 90 120 150 180
HEAOING(DEGREES)

Fi~U~e 2.34. Variation of vertical, horizontal and combined moments

with heading in short-crested seas.

83

93
 :2

v
I I 1
0[
() 10 20 30 40
SIGWICMJ1 WAVC MEICI-IT [FT)

Figure 2.35. Comparison of combined bending moment responses as

computed by three methods.

84
2.1 Rice, S.0., Mathematical Analysis of Random Noise, Bell
System Technical Journal, Vol. 23, 1944; Vol. 24, 1945.
Reprinted in N. Wax, Selected Papers on Noise and Stochastic
Processes, Dover, New York, 1954.

2.2 Crandall, S. and Mark, W., Random Vibration in Mechanical -

Systems, Academic Press, 1963.

2.3 Davenport, W.B. and Root, W.L., An Introduction to the Theory

of Random Signals and Noise, McGraw-Hill, New York, 1958.

2.4 Pierson, W.J. and Moskowitz, L., A Proposed Spectral Form for
Fully Developed Wind Seas Based on the Similarity Theory of
S.A. Kitaigorodiskii, Journal of Geophysical Research, Vol. 69,
December 1964.

2.5 Bretschneider, C.L., Wave Variability and Wave Spectra for

Wind-Generated Gravity Waves, Beach Erosion Board, U.S.
Army Corps of Engineers, Technical Memorandum No. 118,
1952.

2.6 International Ship Structures Congress, Report of Committee 10,

Vol. 2, 1967.

2.7 Cartwright, D.E. and Longuet-Higgins, M. S., The Statistical

Distribution of the Maxima of a Random Function, Proc. R. Sot.
of London, Ser. A 237, pp. 212-232, 1956.

2.8 St. Denis and Pierson, On the Motion of Ships in Confused

Seas,Trans. SNAME, 1955, p. 280.

2.9 Salvesen, N., Tuck, E. and Faltinsen, O., Ship Motions and Sea
Loads, Trans. SNAME, 1970, P.250.

85
 2.10 Abkowitz, M., Stability and Motion Control of Ocean Vehicles,
M.I.T. Press, Cambridge Mass., 1969.

2.11 Korvin-Kronkovsky, B., Theory of Seakeeping, SNAME, New

York, 1961.

2.12 Jacobs, W. The Analytical Calculation of Ship Bending Moment

in Regular Waves, Journal of Ship Research, Vol. 2, No. 1, June
1958.

2.13 Tasai, F., Ship Motions in Beams Seas, Research Institute for
Applied Mechanics, Vol. XIII, No. 45, 1965.

2.14 Tasai, F., On the Swaying, Yawing and Rolling Motions of Ships
in Oblique Waves, International Shipbuilding Progress, Vol. 14,
No. 153, 1967.

2.15 Smith, W. E., Computation of Pitch and Heaving Motions for

Arbitrary Ship Forms, International Shipbuilding Progress,
Vol. 14, No. 155, 1967.

2.16 Raff, A. I., Program SCORES - Ship Structural Response in

Waves. Ship Structures Committee Report SSC-320, 1972.

2.17 SPRINGSEA II - Springing and Seakeeping (Program).

American Bureau of Shipping, New York. Also see SPRINGSEA
H, Program Listing and Users Manual, Mansour Engineering,
Inc., Berkeley, California, 1974.

2.18 Loukakis, T., Computer-Aided Prediction of Seakeeping

Performance in Ship - Design. M.I.T. Department of Ocean
Engineering, Report No. 70-3, August 1970.

2.19 Meyers, W. G., Sheridan, D.J. and Salvesen, N., Manual -NSRDC
Ship Motion and Seaload Computer Program. NSRDC Report No.
3376, February 1975.

86
2.20 Wahab, R., Amidships Forces and Moments an a CB = 0.8 Series
60 in Waves from Various Directions, Netherlands Ship
Research Center TND Report No. 1065, 1967.

2.21 Dinsenbacher, A. and Andrew, J. Vertical and Transverse

Loads and Motions of a Segmented Model in Regular Waves,
NSRDC Report No. 3151, 1969.

2.22 Lewis, E., Predicting Long-Term Distributions of Wave-Induced

Bending Moment on Ship Hu1ls, Spring Meeting, SNAME, 1967.

2.23 Fukuda, J., Long-Term Predictions of Wave Bending Moment,

Part I and Part II, Journal of the Society of Naval Architects of
Japan, Vol. 120, 1966; Vol. 123, 1968.

2.24 Hoffman, D. and Lewis, E., Analysis and Interpretation of Full

Scale Data on Midship Bending Stresses of Dry Cargo Ships,
Ship Structure Committee, Report SSC-196, June 1969.

2.25 Bennet, R., Determination of Wave Bending Moment for Ship

Design. - Paper - presented to the Scandinavian Ship Technical
Conference, Finland, 1964.
9

2.26 Band, E., Long Term Trends of Hull bending Moments,

American Bureau of Shipping, 1966.

2.27 Mansour, A.E., Extreme Value Distributions of Wave Loads and

their Application to Marine Structures. Marine structural
Reliability Symposium, Arlington, Virginia, October 1987.

2.28 Ang, H.-S. and Tang, W.H., Probability Concepts in Engineering

Planning and Design, Vol. II, John Wiley and Sons, 1984.

2.29 Cramer, H., Mathematical Methods of Statistics. Princeton

University Press, 1946.

87
 2.30 Vanmarcke, E.H., On the Distribution of the First-Passage Time
for Normal Stationary Random Process, Journal of Applied
Mechanics, Trans. ASME, March 1975, pp. 215-220.

2.31 Gumbel, E., Statistics of Extremes, Columbia University Press,

New York, 1958.

2.32 Silveria, W.A, and Brillinger, D.R., On Maximum Wave Heights

of Severe Seas, Proceedings of the offshore Technology
Conference, paper #3232, Houston 1978.

2.33 Mansour, A.E., A Note on the Extreme Wave Load and the
Associated Probability of Failure, Journal of Ship Research, Vol.
30, no. 2, June 1986, pp. 123-126.

2.34 Bergman, L.E., Ex~eme Statistics, Risk, and Reliability, Report,

University of Wyoming.

2.35 Little, R. S., Lewis, E.V. and Bailey, F. C., A Statistical Studv of
Wave Induced Bending Moments on Large Oceangoing Tanker
and Bulk Carriers, Trans. SNAME, 1971.

2.36 Critchfield, M., Evaluation of Hull Vibratory (Springing)

Response of Great Lakes Ore Carrier M/V Stewart J. Cort, DTMB
Report 4225, Nov. 1973.

.
2.37 Stiansen, S.G., Mansour, A. E., Jan, H.Y. and Thayamballi, A.,
Reliability Methods in Ship Structures, The Naval Architect,
Journal of RINA, July 1980.

2.38 Stiansen, S.G., Mansour, A. E., Ship Primary Strength Based on

Statistical Data Analysis, Trans. SNAME, 1975.

88
2.39 Stiansen, S.G., Mansour, A.E. and Chen, Y.N., Dynamic Response
of Large Great Lakes Carriers to Wave-Excited Loads, Trans.
SNAME, 1978.

2.40 Ferro, G. and Mansour, A.,

Probabilistic Analysis of the
Combined Slamming and Wave-Induced Responses, Journal of
Ship Research, Vol. 29, No. 3, Sept. 1985, p.170.

89
3. STRENGTH INFORMATION REQUIRED FOR RELIABILITY
ANALYSIS OF MARINE STRUCTURES:

3.1 b tre n tih Va riabilitv and Modelli~

Because of limitation on control of properties of steel and other

materials used in marine structures and because of limitations on
production and fabrication of their components, the strength of apparently
identical marine structures will not be, in general, identical. In addition,
uncertainties associated with residual stresses arising from welding, the
presence of small holes, etc. may affect the strength of the marine
structure. These limitations and uncertainties indicate that a certain
variability in strength or hull capacity about some mean value will result.
This will in turn introduce an element of uncertainty as b what is the
actual strength of the marine stmcture that should be compared with the
loads obtined in the pretious section.

Additional uncertainties in the strength will arise due to

uncertainties associated with the assumptions and methods of analysis
used to calculate the strength. Further uncertainties are associated with
possible numerical errors in the analysis. These emors may accumulate in
one direction or possibly, tend to cancel each other. Whatever the case may
be, the above uncertainties have to be reflected in any reliability or failure
analysis.

Designers and naval architects are aware of the presence of these

uncertainties. However, they are usually treated in a qualitative sense and
very few attempts have been made to quantify them. The qualitative
assessment of the uncertainties does not lend itself to systematic
improvement of design procedures based on previous experience. Full
advantage of that experience can be obtained by attempting to quantify
them.

It is convenient to divide and recognize two types of uncertainties

(hg [3.1, 3.2]):

90

po
1. Objective uncertainties. These are uncertainties associated
with random variables for which statistical data can be collected and
examined. They can be quantified by a coefficient of variation
derived from available statistical information. The variability in the
yield strength of steel is an example.

2. Subjective uncertainties. These are uncertainties associated

with the lack of information and knowledge. They can be
determined only on the basis of the engineers previous experience
and judgment. Examples of these include assumptions of the
analysis, error in the design model, and empirical formulas.

Variability in failure load that will cause yielding of a cross

section or buckling of plating results from uncertainties in the
following factors:

1. Uncertainties associated with the material yield strength and

Youngs modulus of elasticity of different components of the section
such as plates, girders and stiffeners.

2. Uncertainties associated with scantlings of components such as

plate thicknesses, stiffeners, girders, and face plate dimensions.

3. Uncertainties associated with the distribution of residual

stresses due to welding.

4. Uncertainties associated with major dimensions such as the

beam and depth of a cross section.

5. Uncertainties associated with manufacturing imperfections,

flaws, plate fairness, etc.

In addition to the above objective uncertainties, the following

subjective uncertainties cause a variability in the strength or
capacity of the marine structure:

91
1 Uncertainties associated with the degree of effectiveness of plating
due to shear lag effects [3.3, 3.4].

2 Uncertainties associated with the usual Navier hypothesis of plane

section remain plane and perpendicular to the neutral axis
(modelling assumptions).

3. Uncertainties related to the presence of small holes and cutouts that

may etist in the deck plating.

4. Uncertainties associated with the residual strength after buckling

[3.5] and the effect of initial deformation on the buckling loads [3.6].

As more information and more knowledge is accumulated, some of

the factors identified under subjective uncertainties can be classified under
objective uncertainties.

Other classification of uncertainties are also possible and will be

discussed later.

A physical reasoning may be used in the choice of strength or

capacity distribution. Two limiting cases are widely used to represent the
strength of a marine stmcture:

1. The Gaussian (Normal) Distribution:

The central-limit theorem is often used to justify the use of the

normal distribution. It is known that under very general conditions this
distribution arises (or is approached asymptotically) in many
practical problems. A sufficient condition for this is the possibility
of considering the deviation of the given random variable from its mean
value as the sum of a large number of independent random variables
with different probability laws but none of which has large
variance compared with the others. This condition is satisfied in random

92

@
 quantities such as resistance of materials, weight of materials, and
geometric parameters of a section.

Thus in general the strength distribution of any structure

whose strength is a linear function of a number of independent
random variables may be considered to approach the normal
distribution. The rate at which the sum tends to normality depends
on the presence of dominant non-normal component [3.7]. The
normal strength model will be adopted in this report.

2. The Lognormal Distribution:

The lognormal distribution arises as a limiting distribution

when a random variable is a product of a number of independent
and identically distributed random variables.
In modelling the
strength of a component by Iognormal distribution, the advantage of
precluding non positive values is obtained. However, the strength or
resistance of the member should be regarded as the product of a
number of random variables.

3. z Limit States Associated with Marine Structures:

,-

Several limit ktates may be defined for a marine structure.

These include:

1 Ultimate strength limit state (extreme load)

2. Fatigue limit state
3. Serviceability limit state

The ultimate strength limit state can be further decomposed

into two modes of failure:

a. Failure due to spread of plastic deformation, as can be

predicted by plastic limit analysis and fully plastic moment for
beams (initial yield and shake down moments can be also classified
under this category) [3.26].

93
 b. Failure due to instability or buckling of a panel longitudinal
stiffeners (flexural or tripping) or overall buckling of transverse and
longitudinal stiffeners of a grillage.

Each of the above two modes require separate methods o f

analysis and are discussed thoroughly in reference [3.25, 3.26 and
3.8] for ships.

The fatigue limit state is associated with the damaging effect of

repeated loading which may lead to a loss of a specific function or to
ultimate collapse. This particular limit state requires independent
type of analysis and is treated in a reliability framework in
chapter 9.

The serviceability limit state is associated with constraints on

the marine structure in terms of functional requirements such as
maximum deflection of a member or critical buckling loads that
cause elastic buckling of a plate.

3.3 Analysis of uncertainty:

As discussed earlier uncertainties can also be classiiled under

objective and subjective uncertainties. They can also be classified
under inherent and model uncertainties. The former is associated
with physical phenomena that are inherently random such as height
of ocean waves. The latter is associated with models for estimation
or prediction of reality such as theoretical models for predicting
ultimate strength of a marine structure or imperfection of sampling
of yield stress tests.

Uncertainties, both inherent and model, can be expressed in

terms of a full probability distribution or more simply by coefficient
of variation (c.o.v.). The method of quantification depends on the
form of available data [3.9].
a. Sample Data Available [3.9].

Consider a set of sample data ( Xl , q, . . . . . . ,x. ). The mean ~ ancl

standard deviation &Xcan be estimated from
n
2= 1 x.
n z 1
and i=l
n
2 1
a = ( x. -Z)a
x n- 1 E 1
i=l
thus the inherent uncertainty as given by a C.O.V. is
u

= x
6
x x
The estimated mean value may not be totally accurate in
comparison to the true mean value because of the sample size n. The
sampling or model emor in estimating ~ is

Therefore the uncertainty associated with the sampling (model)

error is
v-

Ax=
;

b. Range of Values Known

In estimating uncertainties that require judgement, it is often

convenient and mom realistic to express each in the form of a range,
i.e., upper and lower bounds (e.g. ultimate strength of a member).
Consider now a random variable X with lower and upper bounds
%! and %U , respectively. The mean and C.O.V. can be determined
depending on the assumed distribution. Ang and Tang in reference
[3.9] give these values for several representative distributions as
follows:

If a uniform distribution is prescribed between X ~ and K. then

 1
z= 2
(X1+X)

and x -x
6
=*(X:+X:)
f(x)

m.
u-xl

xl x~

Figure 3.1. Uniform p.d.f.

If a symmetric triangle distribution is assumed, then

2= +(X1+X)

f(x
I

2

Xu-xl

u
Figure 3.2. Symmetric Triangular p.d.f.

96

[oL
 If there is a bias towards higher values then the upper triangle
distribution shown in Fig. 3,3 is more appropriate, and in this case

z= +(X1+ 2x
u)
x x
6Y=A(
F 2XU
+X: )

f(xJ
Lower
Triangle

J
1 \
:
I
x
xl %
Figure 3.3. Upper and Lower Triangle p.d.f.

If there is a bias towards the lower range, then a lower

triangular distribution may be used (see Fig. 3.3), thus,

and
x -x
8
+:++ )
If a normal distribution is assumed (Fig. 3.4) with specified
limits of ~ k standard deviation, then

1
%= ( l.+X )
2

and
x-x
u
6X
-& ( ~ U+x
1
1
)

97

[01
 f(x)

x
xl x~

Figure 3.4. Normal Distribution.

3.4 Random Error Analysis:

In the calculation of the strength parameters, use is usually

made of the theory of error. This theory can be found in applied
statistics books such as [3.10], and its application to ships has been
discussed in [3.11] and to other structures [3.12, papers (i), (ii) and
(iii)]; therefore, it will not be repeated here. It is shown in
references [3. 10] and [3. 11] that if the strength, S, has a functional
relationship with its constituent partss , & ~El -. -.~n , in the form

tn) (3.1)

and if t; are independent, then, the approximate estimate of the

mean ~ ,and variance ~zof S are given by9 :

. . F*)
(3.2)

8 ~; are the random variables which affect the strength S such as

yield stress, plate thickness, stiffeners dimensions, etc.

9 The random variables, Ei are assumed to be closely distributed

about their mean; S and its derivatives are assumed to be
continuous.

98
 = k @s/tk,)t
i-l
g,,z

(3.3)

where ~: and ~fiAare the mean and the variance of the random variable
E: The p~tial derivatives in equation (3.3) are to be evaluated
at the mean value Et . Equation (3.3) can be normalized and written
in terms of the coefficient of variations (COV):

(3.4)

where 6%= ~ is the strength COV, and s~i=~: /~L are the COV of &i and
the partial derivatives are to be evaluated at the mean values.

If the variables ki are correlated and the correlation

coefficients ~ij between EL and ~~ are known, then equation (3.2)
still gives the mean of the strength S in terms of the means of ~1 but
the C.O.V.of $ becomes

6== z +
P
-

))* .-
. P.
as
lJ ( aci )(*)UU
J
E.
1 J (3.5) .
i J

where the partial derivatives are evaluated at the mean values.

3.5 Uncertainties associated with ship strength:

The strength or a limit state associated with a ship is a function

of several variables. In order to determine the mean and C.O.V. of the
strength, information must be obtained on those variables affecting
it. With that purpose in mind the following variables were
evaluated: material yield strength, material ultimate strength,

99
Youngs modulus, ship steel plate thickness, ship steel corrosion
rates, residual stresses, and fabrication tolerance.

Computerized on-line periodical searching was used [3. 13] to

cover efficiently as much ground as possible for three of the
variables. Several data bases, available from the DIALOG system and
representing several million citations of journal articles, symposia,
conference papers and U.S. Government technical reports, were
searched. This yielded about 300 citations which were further
reviewed for suitability and included as appropriate. The striking
result from these broadly based searches was the lack of statistical
data on the appropriate variables, even though extensive literature
exists on the subjects in general.

Yield stren~th. ultimate strength, and Youn#s modulu~ The

measurement of yield and ultimate strength is the most basic test
that can be made in materials research, yet most papers will include
only one or two tests to show the relative merits of a new process or
alloy. Even when statistics are reported, another difficulty arises,
namely, lack of uniformity in the test method and results reported.
Under the general category of yielding, there are measurements of
proportional limit, elastic limit, yield strength (0.2 percent off set),
yield stress level, upper yield point, and lower yield point. Alpsten
[3. 14] discusses the weakness of each one of these measurements
and points out that all of them are affected by the strain rate and or
residual stresses in the sample? He recommends use of the 0.2
percent offset measurement because it compares well with static
strain tests. Measurement of yield points is particularly sensitive to
strain rate, but most of the older data do not report this value.

Table 1 provides a summary of more than 60,000 samples of

various steel types and test methods. Galambos [3.15] in reviewing
much the same data suggests that any numerical analysis is probably
worthless since the measurements are so varied. His judgment is
that for rolled shapes the mean yield stress be taken as 1.05 Fti in
flanges and 1.10 ~a in webs with CC)VS of 0.10 and 0.11, respectively.

100
Fj is the specified yield stress for the steel grade used. The
weighted average of the COVS for the data presented in Table 3.1 is
0.089.

The results for ultimate strength are safer to compare since

this measurement is not particularly affected by strain rate. Table
3.2 presents results for about 4200 samples but representing several
different types of steel. Here the weighted average of the COVS is
0.068.

Finally, Table 3,3 gives the results of 300 samples measuring

Youngs modulus. Overall, the weighted average of the mean value is
30.07 x 10 ksi and the weighted average of the COVS is 0.031.

Ship steel plate dimensions: A careful literature review

revealed only limited statistical data on the thickness of ship steel
plate [3.16]. There is an extensive body of literature on the detail of
manufacturing plates and how to improve the quality, but no specific
numbers as to the variations that typically occur. Informal contacts
with one major steel producer tend to confirm that whatever data
are collected in the mills are considered proprietary.

In [3.17] Basar did obtain such data from one manufacturer,

but he did not report any statistics.

~Q; This is another topic where there is a lot of

literature, but few statistical data. Only one paper [3.18] giving
actual shipboard corrosion rates has been found, but the data are
based on observations of a single tanker only. Two other papers ~
were found which have statistics on rates, but they were for
unprotected steel samples [3. 19, 3.20], thus any comparisons are of
limited value. Table 3.4 presents a summary of this information.

Residual stress: Statistical data on residual stress are

extremely difficult to find; in fact, this search revealed very little.
While this subject has received considerable study, the testing

101

H1
method is prohibitively expensive, which precludes gathering
statistically significant amounts of data. Alpsten [3.14], for example,
reports that surveying a single plate for residual stress took 140
hours, not man-hours, but the time it took a team to collect the data.
Two representative papers which give results and empirical methods
for predicting residual stress are included in the list of references
[3.21, 3.22].

~ While no extensive
search was conducted on this subject, it seems safe to say that the
literature on the subject is limited. Basar [3.17] in his survey of
structural tolerances in the U.S. shipbuilding industry, states that

The quantity of structural deviations data obtained was rather

limited partly due to the fact that the yards did not maintain a
statistical record and partly due to the fact that actual measurements
proved to be difficult to carry out in that it interfered with the yards
work in progress.

As far as in service deviations are concerned, again not

enough data were available due to the fact that such data are not
being recorded and sometimes not even reported.

This led the International Ship Structures Congress in their

1976 report to recommend for future research the establishment of a
comprehensive Damage Recording System. The report cites the
need for all parties concerned-that is, the classification societies, ship
owners, and ship repairers-to take a more liberal view of the
subject and to release information of this type for the benefit of the
industry.

Two other related papers were located: [3.23], which gives

statistical data and distributions for steel-plated highway bridges,
and [3.24], which discusses fabrication errors in a Japanese high-rise
building.

102

u-
 Tablcg.d Ylald alrw@h data

No. of Me-n,
Test Samplcy psi Cov [)istribu;ion RHIIirrkh

~icld stress Mw 0.091 resumedhKnortd mill test 1 con~inmcnl vwi SA537 CrR
} iekl slrw :1964)0 0.10 eumcd Iutinormil mill tCSt 1 cmrtainmenl vessel SA36 :
Yield strem :15W1 0.042 extreme Type I cold straightenedsha~ HE~B
Yield stress
J 36583 0.0922 extremeTyIM I coldstrai~htenedshansHE$KW)B
.-
Yield she~ 19857 40073 0.103 extremeType I mill lmta
~itld ~trens 19217 43475 0.099 xtreme Type I mill laatn
\ irld strew 11170 57fr16 0.057 rxtreme Type I mill teats
Yield stress 2447 63.336 0.054 extreme Type I mill tsstn
upper ... mill tuts
yield poim :{974 40m 0.087 hrgnormal impi~ log nnrmal. refers tn Freudenthal,
AS(X, vol. 121, 1956
Yield strm 400 44mm 0.11 ...
Yield stress 33 36091 0.059 .. . 1948 tests, AM Class A platm 71$and VI in,
Yield pnint 79 34782 0.116 .. . 1948 tasta. ABS Class B pks /16.5/u,1l/l& 3/4,
% k %6,1h.
Yield puint 33831 0.081 1948 tests, ABS Class C pkk 1$E,13/ISi1/4,
1%. 11/1in.
Yield strength 39 34850 0.044 normal Yj-in. ABS B Steel
Yield strength 36 35(WI o.(mg .,, I /,.in. Af3S C Steel
Yield point :1124 39:160 o.07n . ASTM mill tests
Yield point 35 42900 ().10:{ mill hmt i - 1(XKIpin./inJscc
Yield .,.
S1r?mIrvd 35 41X4-I 0.1m snmt specim?ns w nlriIvr. wmIIliIled mill lPSI
Ywld strew+Ievd 35 :14m 0.121 wme SfMClmCllSas abe, stub dUInn lest.
stitic
strainraw
Yield stress 0.2% 50 61780 0.034 not normal cold-rolled
rmle
not lognormal
Yield stress0.2% 38 40530 0.045 not normal hot-roildrds
not ~gnormsl
Lnwer
yield 22 30923 0.110 .. . arbon structural ?4- nd I in. plata and /s-in.

Lower yield 20 42675 0.079 ... loflfi$y +4- and l-in. plata
bwer yield 10 M 290 0.068 ... low-alloy Yt.in. late
Yield point 120 35 OEO 0.038 ... ASTM A7-55T $JF barns, flanges
Yield point 39079 o.od4 .. . ASTM
A7-55T, WF hama, web
Yield point 38(WI 0.0261 .. . ASTMA7-55T,WF bea~, cover platea
Yield stre~ 41800 0.10 lo~normal Mill tek 4 different contimmcnt VISAS,
assumed SA516 GR70
Yield stress 0.066 Iognormal Mill tut. 1 con~inment VHS4 SA516GR70
assumed

P
*

. .
Tabla 3.lUtllmalo alrangthdda

No. of
T-1 Samples Mean, kai COV Distribution Remarks
Tcnaion 58.291 0.043 ,.. cold straighten shapa
Tens;on 3; 57.909 0.089 ,.. cold straighten shape
Tension 84.039 0.1124 . annealed, alloy Wed
Tension ; 124.9 0.1796 ... quenched, alloy steel
Ten~ion 60.405 0.0719 ,.. nominal maximum straas, varioua platex. st~ctural skel
Tcnaion ;: 73.525 0.074 . nominal maximum atreca, various platac, low-alloy s*1
Tension 80.39 0.109 ... nominal maximum stress, various plates, low-alloy 9tccl
Tension l;: 62M 0.0226 ... ASTMA7-MT,WF Iwama, flatrgm
Tension 64.33 0.0341 ,., ASTM A745T, WF barns, web
Temion u 60.84 0.0241 ... ASTMA7-55T,cover plates
Tension 39:; 64.27 0.0703 ... mill Wta
Tensile strength 59.27 0.044 ... 1W8testsABSClass A plataq h % in.
Tensile s~ength 79 0.091 ... 1948 tasta ABS Class B plataa, /1~ % 11/1~,~4,1%6,7A, 1%6, 1 in.
Tensile strength 13 R% 0.051 .,. 1948tsaraABSClass~ plata, 1Ylt. l?I~, 1~, 1%,11Ain,
Tension 6257 0.044 norm41
Tensile
strength % o.&7 normal ~;#A&B&&,

103
 Ta&+a 3.3Ywtgs
modulus
dints

1N04
i}f
T=t Samples Mean, ksi Cov Distribution Remarks
Tcnaion 104 30.0x lW O.owln ... structural a~l from brid es
Tension 19 26.98x 102 0.0M9 ... variousM.A lloys,annm! ad and quenched,nd
drawn samplm
Tension 22 29.50xl@ o.(w2 ... structural steel
Compression 29.49x 1(P 0.0146 ... structural akl, same ssmpla as above
Tension :: 29.59x l@ 0.CN156 ... low alloy
ComprSion 20 29.Mx ltP 0.IJ370 ... low alloy, same sampl~ as abve
Tension 10 29.56x l@ O.(HH ... low Slloy
Compression 29.61X 10s 0.01106 ... low 4UOY,same aamplm u ake
Tension $ 29.42X I@ 0.01% ... variom-sixe a~ens from Y4,~ and 1in. plate,
sune WI& atn.rctural 3W1
Ten~ and stindard 94 31.20 x 10s 0.060 ... sttucturd steal
..-. .. -

(i) Mean corrmion ra~ for ~ ~nker

L-5.l&3
CorrosionMean
and .%ndard
Deviation
Membr or Groupi@ miIs/year
herd steel, upper 15 ft 6.5 & 0.4
Internal steel, lower 30 ft 3.3 * 0.2
~k longitudinal 6.5&0.4
Bottom ]ongitudinals 3.3 i 0.2
Deck phe 11.4+0.7
Shell plate sid~ 5.4* 0.6
Bottom phe. wing Mnks 5.4i 1.6
Bottom plate, cen~r bsnka 17.9+ 2.9

(ii) Corrosion Ratesfor Unpro~d Stael in Brackish Water [3 .20J

Mean and Standard

Deviatirmin
Maximum p]t depth
Uniform corrosion

(iii) Corrmion Rat= for Unprotected St4 in Seawater [5. 131

Exrmaure depth 0.5 m 1.5m

Uniform Corrosion. mii/year~ 5.5 * 0.3 5.7* 0.2

Originally reporti as millimetam of corrosi(m after :18
months.
4 originally reported M g/m2 weight ha after 16 mwlths.

N-1
 REmmcEs

3.1 Ang, A. H.-S., and Ellingwood, B. R., Critical Analysis of

Reliability Principles Relative to Design, prepared for the
conference on applications of statistics and probability to soil
and structural engineering, Hong Kong, Sept. 1971

3.2 Ang, A. H.-S., and Amin, M., Formulation of Wind-resistant

Design Based on Acceptable Risk, prepared for the Third
International Conference on Wind Effects, Tokyo, Japan, Sept.
1971.

3.3 Schade, H. A., The Effective Breadth Concept in Ship-Structure

Design, Trans. SNAME, Vol. 61, 1953, pp. 410-430.

3.4 Mansour, A. E., Effective Flange Breadth of Stiffened Plates

Under Axial Tensile Load or Uniform Bending Moment, Journal
of Ship Research, Vol. 14, No. 1, March 1970.

3.5 Mansour, A.E., On the Nonlinear Theory of Orthotropic Plates,

Journal of Ship Research, Vol. 15, No. 4, Dec. 1971.

3.6 Mansour, A.E., Post-buckling Behavior of Stiffened Plates with

Small Initial Curvature under Combined Loads, International
Shipbuilding Progress, Vol. 18, No. 202, June 1971.

3.7 Thoft-Christensen, P. and Baker, M.J., Structural Reliability

Theory and its Applications, Springer-Verlag Berlin Heidelberg,
New York, 1982.

3.8 Mansour, A.E., Probability Design Concepts in Ship Structural

Safety and Reliability, Trans. SNAME, Vol. 80, 1972.

3.9 Ang, A. H.-S., and Tang, W. H., Probability Concepts in

Engineering Planning and Design, Vol. II, John Wiley & Sons,
New York, 1984.

105
3.10 Pardine, C. and Privett, B,, Statistical Methods for Technologists,
the English University Press, London, 1964.

3.11 Mansour, A.E. and Faulkner, D.,On Applying the Statistical

Approach to Extreme Sea Loads and Ship Strength,: Trans.
RINA, Vol. 115, 1973.

3.12 A Statistical Method of Design of Building Structures,

translation by D.E. Allen of five papers in Russian, National
Research Council of Canada, Technical Translation no. 1368,
Ottawa, 1969.

3.13 Mansour, A.E. et al, Implementation of Reliability Methods to

Marine Structures, Trans. SNAME, 1984.

3.14 Alpsten, G. A., Variations in Mechanical and Cross-Sectional

Properties of Steel, Tall Building Criteria and Loading, Vol. Ib,
Proceedings, International Conference on Planning and Design
of Tall Buildings, Lehigh University, Bethlehem, PA, Aug. 1972.

3.15 Galambos, T. and Ravindra, M.,Properties of Steel for Use in

LRFD, Journal of the Structural Division, ASCE, Vol. 104, No.
ST9, Sept. 1978, 1459-1468.

3.16 Staugaitis, C. L., Mill Sampling Techniques for the Quality

Determination of Ship Steel Plate, Ship Structure Committee
Report No. 141, 1962.

3.17 Basar, N.S. and Stanley, R.F., Survey of Structural Tolerances in

the United States Commercial Shipbuilding Industry, Ship
Structure Committee Report No. 273, 1978.

3.18 Purlee, E. L., Leyland, W. A., Jr., and McPherson, W.E., Economic
Analysis of Tank Coatings for Tankers in Clean Service, Marine
Technology, Vol. 2, No. 4, Oct. 1965.

106

[(L
3.19 Josefsson, A. and Lounamaa, K., comparative Comosion Tests
on Steel Plates Rolled from Continuously Cast Slabs and Rolled
from Mold Cast Ingots, Corrosion, Vol. 26, No. 5, May 1981.

3.20 Sabelstrom, S. and Enestrom, L.E., on the Effect of Casting

Method on the Corrosion of Carbon Steel in Water,
Scandinavian Journal of Metallurgy, 1976.

3.21 Bjorhoude, R., Brozzetti, J., Alpsten, G.A., and Tall, L., Residual
Stresses in Thick Welded Plates, Welding Journal, Vol. 51, No.
8, Aug. 1972.

3.22 Nagaraja, N.R. and Tall, L., Residual Stresses in Welded Plates,
Welding Journal, Vol. 40, No. 10, Oct. 1961.

3.23 Massonet, C., Tolerances in Steel Plated Structures, IABSE

Surveys, S-14/80, Aug. 1980.

3.24 Tomonaga, K., Measured Errors on Fabrication of Kasumigaseki

Building, Tall Building Criteria and Loading, Vol. Ib,
Proceedings, International Conference on Planning and
Designing of Tall Buildings, Lehigh University, Bethlehem, PA,
Aug. 1972.

3.25 Faulkner, O., A Review of Effective Plating for Use in the

Analysis of Stiffened Plating in Bending and Compression,
Journal of Ship Research, March 1975.

3.26 Caldwell, J. B., Ultimate Longitudinal Strength, Transactions,

Royal Institution of Naval Architects, Vol. 107, 1965.

107
4. BASIC RELIABILITY CONCEPTS BASED ON FULLY PROBABILISTIC
METHODS - LEVEL 3:

4,1 Introduction n - Reliability Levels;

Structural reliability is currently categorized under three different levels

that depend mainly on the degree of sophistication of the analysis and the
available input information. Level 3, which sometimes is referred to as the fully
probabilistic approach, is the most demanding in terms of the required input
information. But even if the input information is available, the analytical or
numerical evaluation of the restilting integrals for estimating the probabilities of
structural failure is extremely difficult. The basic concept of Level 3 reliability
analysis is that a probability of failure of a structure always exists and may be
calculated by integrating the joint probability density fimction (j.p.d.f.) of variable
involved in the load and strength of the structure. The domain of integration is
over the unsafe region of the variables.

Because of the difficulties in connection with determining the j.p.d.f. of the

variables and in evaluating the resulting multiple inte~ation, Level 2 reliability
(semi-probabilistic approach) analysis was introduced. In this level, a reliability
index, rather than a probabili~ of failure, is introduced to assess the safety of the
structure. The reliability index is c~nne~t~d to the probability of failure, and,
,
under certain circumstances, the exact probability of failure may be directly
obtained if the safety index is determined. For example, if the design variables
are uncorrelated and normally distributed and the performance function 10 is
linear, the probability of failure can be determined from the safety index using
tables of the standard normal distribution function. If the variables are
correlated and not normally distributed, certain transformation

10 The performance function is a function that contains the load and

strength variables and determines the performance or the state of the
structure.

108
 (Rosenblatt transformation (1 969)) can be made to obtain equivalent
uncorrelated normal variables, thus, approximate probability of
failure may be determined. Similarly, certain approximation can be
made for nonlinear performance functions.

Originally, the safety index (Level 2) method was based on a

simple mean value first order second moment analysis (MVFOSM),
see for example [4.1, 4.2, 4.3].

Later Hasofer and Lind [4.4] introduced a more consistent

invariant method based on first order reliability which entails
expanding the performance function in a Taylor series at the most
likely failure point and retaining only the frost order terms.

Although Level 2 is easier to apply in practice, it is still of

limited use to practitioners. Normally a designer needs factors of
safety to apply in the design process such as those applied to the
yield strength of the material and to the loads. This need resulted in
the introduction of Level 1 reliability analysis. In this level partial
safety factors are determined based on Level 2 reliability analysis.
If these factors are used in a design, their cumulative effect is such
that the resulting des~gn will have a certain reliability level (i.e., a
certain safety index). Thus, code developers and classification
societies may determjne (and specify in their codes) these partial
safety factors that ensure that the resulting design will have a . .
specified reliability level.

Level 3 reliability is discussed in this chapter. The following

two chapters describe Level 2 and 1, respectively.

42 The Basic Problem - Level 3

Level 3 reliability is based -on the direct integration of the joint

probability density function (j.p.d.f.) of the random variables
involved in a design. Therefore it will be called here the Direct-
Integration Method.

109
 The probability of failure or the probability of reaching a
specific limit state is determined from:

pf = J ..*J f&(x, ,x2 *.** Xn) dxldx 2 . . ..dxn

(4.1)

where fx ( . ) is the joint p.d.f. of the important design random

variables X 1 . . . Xn. The domain of integration is over the unsafe
region of a limit state associated with a structure. The limit state
function may be represented as g (XI , x2 . . . xn ) and the
corresponding unsafe region (integration domain of equation (4.1)) is
given by:

fdxl, x 2 . . ..xn) ~~
(4.2)

The above general equations can be simplified for specific

cases. In fact, the first basic reliability analysis started with two
variables only, the strength of a member S and the load acting on it
z. In this case, instead of the n variables described in equation
(4.1) , we have only two variables Xl = S, X2 = Z. Failure occurs when
the load Z exceeds the strength S. The unsafe region is therefore

dxl ,x2) = g(s, z) = s - z 5 0 (4.3)

The probability of failure for statistically independent S and Z is then

given by (s= equation 1.3) :

Pf P[s -z So]=m F~tz) fz(z) dz

Io (4.4)
a

[1 - F+) ] f~(z) dz (4.5)

f
o
These two convolution integrals given by equations (4.4) and
(4.5) can also be determined from the general formulation for n
variables given by equation (4.1). In M case of two variables S and
Z, this equation reduces to:

110
 pf z
I I s,2 (S,2) d= dz
[(s,2):s-250] (4.6)

and if S and Z are independent, then

pf =
I [ f@ fZ(z) ds dz
[(s,2):s-2s0] (4*7)

from which equations (4.4) and (4.5) can be obtained.

Figure 4.1 shows graphically the pdf of S and Z. The overlap of

the two curves represents a qualitative measure of the failure
probability. This figure shows that a reduction of the probability of
failure can be achieved by:

a. increasing the distance between the means of the two

probability density functions fs (s) and fl [%) , that is, by increasing the
mean of the strength or decreasing the mean of the load.

b. decreasing the standard deviation (or c.o.v.) of either p.d.f., that

is, decreasing the uncertainty.

s)

Figure 4.1. Probability of Failure

The reliability of the structure can be measured by the

probability of survival or the non-failure probability given by

111
 P~=l-pf
(4.8)

4.3 The Normal Tail and Margin of Safety:

The probability of failure can be given also in terms of a

margin of safety M defined as the difference between the strength
and the load variables, i.e.,

M= g (X1,X2) =s -z
(4.9)

the probability of failure is therefore

o
Pf = PIM<O] = fM(m) dm FM(0)
f
.m (4.10)

This is represented graphically in Fig. 4.2.

fM(m)
1

area= pf

Figure 4.2. P.d.f. of the Safety Margin.

If both S and Z are normally distributed and independent, then

M is also normally distributed with a mean and standard
Fti
deviation ~~ given by -

(4.11)

and

112
 2 2

M=%+=: (4.12)

The standard distribution of M is obtained by subtracting the

mean ~~ and dividing by the standard deviation i.e. ~~~ . This
standard normal variable has a zero mean and a unit standard

-
deviation, i.e., N(O,l). Equation (4.10) then yields
PM
Pf =FM(0)=#
[1 M
1 - 4 (p)
(4.13)

where PM
$z
M
is called the reliability or safety index, Note that the probability of
failure decreases as the safety index increases.
P

In this particular simple example of independent normally

distributed variables the probability of failure can be exactly
determined from ~ and ~ is given by .

and
P

:=% (4.14)

P= q-~) =1-r(d)
f (4.15)

For example, if ~ = O, Pj= 0.50; if /?= 1,?} =().16 ~dif ~ = 3.1, PF= l(j3
The relations k- 6 and p ~ for other distributionof S and Z with
a mrgin gi~ by equation (4.9) are plottedin section11,1.2of @p31diX 1.

4.4 Probabili ~ of Failure of a Ship Hull Girder

Although level 3 reliability is usually limited in application to

actual structures because of the complexity of the analysis, this level
of analysis can be still applied to assess ship primary strength when
the ship is considered as a beam. In order to determine the
probability of failure as described in the previous sections one has
first to determine the probability distribution of the total load (in

113

623
this case, the total bending moment) acting on the ship.
The total
bending moment consists of stillwater bending moment and the
extreme wave bending moment as developed by one of the methods
described earlier.

In principle, the magnitude of the stillwater bending moment is

also a random variable since it is a function of the cargo distribution
and the shape of the wet envelope which contains a certain amount
of randomness. However, the variability in stillwater bending
moment due to random factors is expected to be much less than that
in the wave moment, and, may for this reason, be considered as a
deterministic quantity of a certain maximum value. This maximum
value may be determined from the distribution of cargo that gives
the maximum permissible stillwater bending moment for the
operation of the ship according to classification society rules.

Actual data on stillwater bending moment analyzed by Soares

and Moan in reference [4.5] show that the normal distribution fits
well the data. In this example we will consider both cases. The
stillwater moment is first considered as deterministic quantity with
maximum value m. and then is considered normally distributed with
mean m and standmd deviation as.

First let Zn = m. + Yn represent the extreme amplitude of the

total bending moment in n- encounters, where m. is the stillwater
bending moment considered to be deterministic. Yn is a random
variable representing the extreme wave bending moment using
order statistics and based on the Weibull distribution with
parameters k and 1 as the initial distribution.

The probability density function and the distribution function

of Z~are given respectively by [4.6]:
nl zm~-~
42.(4 = ~ y
()

-e -(%9.[,- ,-(7)]-1 .> ~ (4.16)

= o otherwise

114
 @zm(,, = [1- ,-(s%=)]9,> ~
. 0 otherwise (4.17)

The above results are extended to the case when the variability of
the stillwater bending moment is not neglected. The stillwater
bending moment is considered to follow a normal probability law
given by

(4.18)

where +T ( b) is the probability density function (pdf) of the stillwater

bending moment T, ~~ is its standard deviation, and m is the mean.

The extreme total bending moment Z ~ is simply the sum of the

stillwater bending moment and the extreme wave bending moment

Zn=T+Yn (4.19)

The distribution function of Z ~ is

,-

. .

(4.20)

where ~YA,r (.9 . ) is the joint probability density function of the

random variables Y ~ and T; and the domain of integration is overall
values of y ~ and t such that yfi +Ef%; t is a dummy variable.

115
Differentiating the last equation of (4.20) with respect to z to get the
pdf of Zn, we get

+s(4 = - #rir(Ylz y) ~Ym

so (4.21)

The stillwater bending moment T is assumed to be statistically

independent of the wave bending moment Yn. Therefore equations
(4.21) and (4.20) can be written respectively in the form

wher,e the dummy variable Yn is changed to y and t to t for

simplicity of notation.

Using the Weibull distribution as an initial distribution for Yn

and equation (4. 18) in (4.22) and (4.23), we obtain

and

d--

The safety R is represented by the ratio of the strength to the

extreme bending moment, R = S/Zn. The probability law of R can be
determined from the probability laws of both S and Zn. If statistical
independence is assumed between the extreme total bending
moment Zn and the strength S, then the probability density and
distribution functions of R are given respectively by

(4.26)

116
 FR(r) =J+. o
42.(4-F~(rz)& (4.27)

Integrating equation (4.27) by parts and noticing that the

probability of failure pf is

P/ - P[R <11- ~R(l)

(4.28)

pf can be written either in the form (see also equations (4.4) and
(4.5))

(4.29)

or
+.
p/=l -
J0 @zm(z)Js(z)dz (4.30)

The strength S is assumed to be normally distributed with

mean = v and standard deviation = a. Its probability density function
is given by

(4.31)

and the corresponding distribution function is

F~(s) -
d
I j~(8)d!l- *=
-- ()s u (4.32)

where Ys ( . ) is the standard normal function tabulated in many

statistics books.

Two cases will be considered next. First, the case when the
stillwater bending moment - is regarded to be deterministic of value
m o. Using equation (4.17) for Ozn(z) and equation (4.31) for fs(z) in
equation (4.30), and noticing that z > m., the probability of failure
can be written in the form

117

121
 A.s
0

p/-l- [1 - ~-(r-ww ]u .,-i(3&

(4.33)

Equation (4.33) is a function of n; i.e., the probability of failure

is dependent on the number of encounters or the length of time the
ship is underway (see Fig. 4.3). It should be noticed that as n + =,

[1 - #r+/w]~ ~oandpt~ 1

that is to say, as the ship encounters wave loads for an infinitely long
time, failure will eventually occur with probability equal to 1.

FREQUENCY
OF OCCURRENCE
f
f
SAGGING HOGGING

P*
LOAD BENDINGMOMENT OR ULTIMATE
) BENDINGSTRENGTH

Figure 4.3. Probability of failure under extreme bending moment

A closed-form solution of equation (4.33) in its general form is

not possible. It is best at this point to specialize in the short- and
long-term analyses individually.

For short term, (Rayleigh initial distribution ) 1 = 2 and k =~, 11)

and assuming, for the moment, n = 1, equation (4.33) becomes

W ~ = ~~ = 2 x man square of the process.

118
For long term exponential distribution, ~=landk =~, and letting
n = 1, equation (4.33) becomes

p,,*d - [1-++)]
u-mo d
+e -T%.*
( P - ~
..
u
@
k ) (4.35)

= Pir+ Pf = Pf

where ~ ( . ) indicates the standard normal distribution function of

(*).

We now return to the second case where the stillwater bending

moment is assumed to be a random variable that follows a normal
distribution with mean m and standard deviation G3 . Substituting in-
equation (4.29) for @Zn from equation (4.24) and F~ (z) from equation
(4.32), the probability of failure in this case is given by

= ;+= , S+-
f-+= Jo @/k) -l

*[(Z - $J/ukiz
(4.36)

For short term, , [ = 2 and k =~, and letting n = 1, equation

(4.36) reduces to

Prl.-l - (1/%)v2/@
s J ~
.0

~ *[(Z - P)/@l - (yi~a

.e-@-:c*m)* &@ (4.37)

For long term, ~ = 1 and k =~, and letting n = 1, equation (4.36)

becomes

119

[27
 w(z:-:)v(
=)& 38)
where equation (4.38) can be further reduced by letting c~ =0 and
equating it to equation (4.35) with m. = m. If the fiist term in (4.35)
representing the probability of failure under stillwater bending
moment is neglected, then

The equality in equation (4.39) holds if

v =mf =1.0
( U* )

in the relevant range of z. This leads to the condition

P m? u; + 4.O(U,+ u) (4.40)

Using equation (4.39) in equation (4.38) yields

For the more general- case when n >1, Bernoulli trials may be
assumed, and the total probability of failure, pf, can be written in the
form

n
Pf=l-[l-pfln=ll (4.42)

120
 Equation (4.42) is applicable for both short and long terms, and
also for deterministic and random stillwater bending moment.

It should be noted that several values may be assumed, in the

case of deterministic stillwater moment, m v , over different periods
of the life of the ship, if a long-term analysis is considered. The
corresponding values of n and the probabilities of failure during
these periods can then be calculated. Since different values of ma
correspond to mutually exclusive events, the total probability of
failure is equal to the sum of probabilities of failure. In this way an
allowance can be made for discrete variation of m ~ .

Although the derivation of the probability of failure for a ship

as given above seems to be complicated, the final results are not.
Equations (4.34) and (4.35) for a deterministic stillwater moment are
simple algebraic equations that can be easily used to calculate p ~ for
a given set of values for the variables. For the case of a random
stillwater bending moment, equation (4.41 ) for the long- term
analysis gives a simple means for calculating p ~ . Equation (4.3 7)
which gives p ~ for a random stillwater moment under stationary
condition (short- term) is not simplified further since these
conditions are- unrealistic from a practical point of view. In the
short-term analysis (p.g. a storm condition) it is more appropriate to
consider a deterxqinistic constant value of the stillwater moment
rather than a random one, therefore, equation (4.34) is more .
appropriate to use. Finally, equation (4.42) may be used to calculate
p~ for values of n larger than one. It should be noted that this
equation may give large errors if n is very large, and, in this case
numerical integration may be necessary. Numerical examples of the
use of the derived equations for p will be given later (-@er 10).
F

Notice that, in principle, the hull girder may fail in hogging or

sagging mode (see figure 4.3). Since these two events are mutually
exclusive events, i.e., the hull can be either in a sagging or a hogging
mode, then the total probability of failure is the sum of the two

121
individual failure probabilities provided that the distributions are
determined separately. In practical application, however, the total
probability of failure is controlled by the direction of the stillwater
bending moment (just as in the deterministic approach). Thus if the
stillwater bending moment is a hogging moment, the total probability
of failure is simply equal to the probability of failure in that mode
since the probability of failure in the sagging mode will generally be
very small. Sum naml vessels ~r, may Ha- fmm this rule.

122

\32
4.1 Cornell, C.A., A Probability-based Structural Code, ACI Journal,
Dec. 1969.

4.2 Mansour, A.E., Approximate Probabilistic Method of Calculating

Ship Longitudinal Strength, Journal of Ship Research SNAME,
Vol. 18, No. 3, Sept. 1974, p. 203.

4.3 Ang, A. H.-S. and Cornell, C,A., Reliability Bases of Structural

Safety and Design, Journal of Structural Division, ASCE, Vol.
100, No. ST9, NOV. 1974.

4.4 Hasofer, A.M. and Lind, N. C., Exact and Invariant Second
Moment code Format, Journal of the Engineering Mechanics
Division, ASCE, Vol. 100, No. EMI, Feb. 1974, p. 111.

4.5 Soares, C.G. and Moan, T., Statistical Analysis of Stillwater

Bending Moments and Shear Forces in Tankers, Ore and Bulk
Carriers, Proceedings, Third Iberoamerican Congress of Naval
Engineering, Madrid, 1982.

4.6 Mansour, A.E., Methods of Computing the Probability of Failure

under Extreme Values of Bending Moment, Journal of Ship
Research, Vol. 16, No. 2, June 1972.

123
 5. LEVEL 2 RELIABILITY ANALYSIS

As mentioned in Chapter IV, Level 3 reliability analysis can be

very difficult to apply in practice. The two main reasons for this is
the lack of information to determine the joint probability density
function of the design variables and the difficulty associated with the
evaluation of the resulting multiple integrals. For these reasons,
Level 2 reliability was developed. In Level 2, the safety index
concept which was first introduced by Cornell [5.1] in 1969, was
further developed by several researchers. In the next few sections,
the development of Level 2 reliability will be presented starting with
the simple safety index concept, followed by several improvements
of the concept.

5.1 The Mean-Value First-Order Second-Moment (MVFOSM) Method :

If Z is a random variable representing the load and S is a

random variable representing the strength, then the safety margin as
defined previously is:

M= S-Z (5.1)

Failure occurs when the total applied load Z exceeds the

ultimate capacity S, that is, when the margin M is negative.
Therefore, the probability of failure pf is

Pf = P[M s O] = FM(0) (5 .2)

For statistically independent load Z and strength S, the mean

~- and variance ~~ of the margin are given by
P-= p-p
m s z
2 2 (5*3)
vm *O 2+a
s z
The standardized margin G, which has a zero mean and a unit
standard deviation, can be written as

124
 M-pm
G=
u (5.4)
m

Failure occurs (or a limit state is reached) when M ~ O so that

equation (5.2) can be written as

Pf = FM (0) = F~
[ ~: 1= G -d) 5-5)

where ~ =~~fn= safety index, which is the inverse of the coefficient of

variation of the safety margin.

If the distribution function F ~ (.) is known, then the exact

probability of failure associated with the safety index can be
determined.12 But even for unknown or unspecified distribution
function FG (.), there will be a corresponding though unspecified
probability of failure for each value of . Thus may be taken as a
P P
safety measure as is the case in the MVFOSM method.

The foregoing results can be generalized as follows. Define a

limit state (or performance) function g(.) as

M=g (Xl,xz ..*. xn) (5.6)

where &~ are the applied and strength parameters considered as

random variables, and the limit state function g(.) is a function that
relates these variables for the limit state of interest (serviceability or
ultimate state). The limit state is reached when:

M= g (X LSX2 ..*. xn) so (5.7)

Notice that the above equation is the same as the integration

domain in the Level 3 reliability (see equation (4.2) ). The limit state
function can be expanded using Taylors series, and if only the first
order terms are retained, we get

12
Se section Al. 2of Appmdix 1 for the relationship h~ @and the
probahili~ of failurefOr severaldistributions.
 *
where z ~ is the linearization point, and the partial derivatives are
evaluated at that point. In the MVFOSM method the linearization
point is set at the mean values ( El, -Zz -- --- ~n ).

The mean and variance of M are then approximated by

Bm=g(zls% . . ..zl) (5*9)

(5.10)

where f!i ~d s the correlation coefficient and the subscripts ~i and

.~~ denote evaluation of the partial derivatives at the mean point.

The accuracy of equations (5.9) and (5. 10) depends on the

effect of neglecting the higher-order terms in equation (5.8).

If the variables ~i are statistically uncorrelated, then (5.9)

remains unchanged but (5. 10) becomes

(5.11)

As an example, if the margin M is represented by the variables

S an Z only, that is

~ = g(x],xd= g(s,z)= s - z

then applying equation (5.9) and (5. 11) for determining the mean
and variance, one immediately obtains identical results as given by
equations (5.3). Equations -(5.2) and (5.5) follow accordingly. This
method is called the ~VFOSM method because the linearization of
the limit state function takes place at the mean value (MV); only the
first-order (FO) terms are retained in Taylor series expansion, and up
to the second moment (SM) of the random variables (means and

126

131)
variances) are used in the reliability measure rather than their full
probability distributions.

A geometric interpretation of the safety margin M = S - Z will

be useful particularly for the discussion of the Hasofer Lind
reliability index which will be presented later. First we notice that
M >0 represents a safe state or region, M < 0 represents failure state
and M = O represents a limit state or failure surface (or line in the
mse of two ti~les). h standard or reduced variates of S and Z can
be written as
s-p Z-pz
s .* 21 =
s = o u
s z

Therefore, the limit state function M = O can be written in the

space of reduced variates as:

M= UgS-UzZ+P~-lJz =0

which is a straight line shown in Fig. (5.1).

, .

Failure
~gion

s
o
Figure 5.1. Limit State Function in the Space of Reduced Variates

The region on one side of the straight line which contains the
origin O represents the safe state (M > O) and the other region
represents the failure state (M ( O). Thus the distance from the
origin to the line M = O can be used as a measure of reliability. In

127

131
fact, from geometry, the minimum distance D shown on Figure 5.1
is given by

Notice that D is equal to the safety index ~ for the case of the
normal variates and linear limit state function discussed earlier, i.e.,
for this case

Pm
$ =D= a
1++
and the probability of failure is thus

Pf = 4 (-D)

5.2 Improvements to the Mean Value First Order Second Moment

Reliability Index:

The MVFOSM method described previously has three basic

shortcomings:

First, if g(.) is nonlinear and the linearization takes place at the

mean values of %; , errors may be introduced at increasing distance
from the linearization points by neglecting higher-order terms.

Second, the method fails to be invariant to different equivalent

formulations of the same problem. In effect this means that the
safety index ~ depends on how the limit state equation is formulated.
For example if the M is set to be a nonlinear function of ,S and Z such
as
M=s2_z2

128
 then PF = FM (O)* still given as before by equation (5.5); however,
when r h and r. are computed from (5.9) and (5. 11) and
substituted in P
m
$ =
0
(5.12)
m
the following is obtained
!
P; - P:
$=
4p: 0; + 4p2Zz 02 05 (5.1s)
[ 1

which is different from the obtained when M is taken as M = S - Z,

P
even though the criterion of failure is still given by equation (5.5).

Third, in the MVFOSM method the safety index ~ can be

related to a probability of failure in cases when the variables xi am
normally distributed [and when the function g(.) is linear in xi]. It is
known that wave bending moments in ships follow a Weibull or
exponential distribution. Thus, one of the improvements in an
advanced method over the MVFOSM method would be, to include
such distribution information (s= -tion 5.2.2).

5.2.1 The Hasofer/Lind Index:

The first two shortcomings discussed previously are avoided by

using a procedure usually attributed to Hasofer and Lind [5.2].
Instead of expanding Taylors series about the mean value point,
which causes the invariance problem, the linearization point is taken
at some point on the failure surface. On the failure surface, the limit
state function g(.) and its derivatives are independent of how the
problem is formulated.

In the Hasofer/Lind procedure, the load and resistance

variables, Xi, are transformed to reduced (standardized) variables
with zero mean and unit variance by

129

13q
 x. - ii.
1 1
Yi =
u
x.
1
(5.14)

The Hasofer/Lind reliability index is defined as the shortest

distance from the origin to the failure surface in the reduced space.
This point is found by solving the following set of equations

(5.15)

(5.16)

(5.17)

G(.) is the
*
failure surface in the reduced space, and , are coordinates of the
Y
point closest to the origin in the reduced spkce (the checking point).
All partial derivatives are evaluated at the checking point. In effect,
this procedure is equivalent to linearizing the llmit state function in
-. the reduced variables space at the checking point and computing
P associated with that point.

In the original space, the checking point or the most likely

failure point is obtained from
*
x.
1
=%.+0
1 x Y:
i
=2 .-0 (5.1s)
1 ai $
i

In general, for a linear limit state function, the Hasofer/Lind

method will yield the same result for ~ as the MVFOSM method. For
nonlinear limit state functions, this method yields a safety index ~ which
is invariant to the formulation of the performance function. TO ill~~~
this @nt, the fdlwtig _le is considerd (seeref. [~. Ill ).

130
 Suppose that a simple beam is subjected to random loading that produces a
maximum stress with a mean value ~Z = 20,000 p.s.i. and a standard deviation Oz
= 2,500 p.s.i. The beam is made of material of mean yield strength P6 = 30,000
p.s.i. and a standard deviation cr~= 3,000 p.s.i.

The following three limit state functions are considered. They all represent
failure of the beam and, therefore, should yield the same value of the safety index
if the method used to determine the safety index is consistent (i.e., invariant to the
formulation of the limit state function). The three limit state functions are:

MI = S-Z (5.19)

Mjj = S2.Z2 (5.20)

w = ~+yz (5.21)

The strength S and the load Z are independent; S is normally distributed

and Z follows a Weibull distribution. The safety index ~ will be computed for each
of these three limit state functions using, first, the mean value first order second
Imoment method (MVI?OSM) then the Hasofar-Lind method. Notice that the first
limit state function (5.19) is linear; the other two are non-linear.

P. The Mean Va lue First Order Second Moment Meth@

In this method the safety index ~ is defined as (see equation 5.12)

Where the margin means ~m and its standard deviation ~m are computed
for each limit state function using equations (5.9) and (5. 11), respectively. The
resulting ~s for the three limit state equations (5. 19) to (5.21) are:

131
 k-vz
h = (5.23)
(0s2 + aza) 1/2

~2.p;
132 = (5.24)
(4ps26s2 + 4p~20.2)1/2

b p@ - hp~
133 = (5.25)
(%% 2 + GZVpz 2 )1/2

Substituting in the above equations the values of ps, ~z, as andaz, one
obtains the following results:

h = 2.5607 ~ = 2.4282 P3 = 2.5329

The values of Ps are not the same indicating that the MVFOSM method is
not invariant to mechanically equivalent formulation of the same problem. Notice
that we have not made use of the distribution information (S: Normal and Z:
Weibull) in the calculation. of the ~ values.

If the probability of failure for each limit state is to computed from pfi = @(-~j,
an emor will result since this equation is valid only if all the random variables are
normally distributed and the limit state fmction is linear. Let us, however, use
this equation in order to compare the results with those obtained by a more
accurate method described in the next section entitled Inclusion of Distribution
Information. Using,

(5.26)

the following values are obtained:

WI = 0.005223 Pf2 = 0.007574 Pf3 = 0.005655

132
b, The Hasofar/Lind Method

Equations (5.15) to (5.17) are applied to determine the ~ value for each limit
state function according to this method. For the non-linear limit state functions,
these equations must be solved iteratively since the evaluation of the derivatives
required for calculating ~ will depend on the coordinates of the most likely failure
point which are unknown. An iterative procedure would be simply to assume
values for the most likely failure point (e.g. mean values of the variables) and to
evaluate the derivatives of the limit state function at that point as required by
equation (5.17). Equation (5.17) is then substituted in (5.16) to obtain a set of
coordinates y? which will be a function of the unknown ~. These coordinates are
substituted in (5.15) and the resulting equation is solved for ~. The obkined ~ is
then used in (5.16) to obtain a new set of coordinates of the most likely failure
point. The procedure is repeated until convergence is obtained. The procedure
will be described in more detail in Chapter 6.

The procedure was applied to the limit state equations given by (5.19) to
(5.21). The linear limit state function (s.19) did not require any iteration; the
second limit state function given by (5.20) required five iterations and the last one
(5.21) required seven. The following results were obtained.

h = 2.5607 - ~ = 2.5607 ~ = 2.5607

These results indicate that the value of ~ is invariant to the formulation of

the problem. They also show that the MVFOSM method gives identical result to
the Hasofar/Lind method if the limit shte function is linear (see ~1 obtained using
the MVFOSM method).

The probability of failure calculated from equation (5.26) for all ~ values
according to Hasofar&ind method is 0.005223. Here again no use is made of the
distribution information given in the problem. Therefore, unless all variables are
normally distributed (which is not the case in this problem), the probability of
failure computed using equation (5.26) will be in error. More accurate values of
the probability of failure will be given for this example fir the next section.

133

w?
5.2.2 Inclusion of Distribution Information:

The third and final refinement of the advanced method over

the MVFOSM method is that the information about the variable
distributions (if known) can be incorporated in the computation of
the safety index and the corresponding probability of failure. In the
MVFOSM method the index ~ can be related to the probability of
failure in cases when the variabIes ~~ are normally distributed [and
when g(.) is linear in Ki ]. This relation is given by [see equation (5.5)]

(5.27 )

where ~ is the standard normal distribution function. In order to be

able to use equation (5.18), approximately, in the case of non-normal
variables, a transformation of these variables into equivalent normal
variables is necessary prior to each iteration in the solution of
equations (5.1 5) to (5.17).

The tail of the distribution is usually the location where most of

the contribution to the probability of failure comes from. It is,
therefore, better to fit the normal distribution to the tail of the non-
normal distribution at the linearization point Z?, which is where
failure is most likely to occur (that is, minimum ~ ). This is the basic
idea of the Rackwitz/Fiessler method as given in [5.3].

By requiring that the cumulative distributions and the

probability density functions of both the actual distribution and the
normal distribution be equal at the linearization point, one can
determine tie man Px and standard deviation a ~ of W equivalent no-
variable, that is

(5.28)

and

134
 M%)
=* ++7-71
(5.29 )

where + (.) is the standard normal probability density.

Since we are concerned with finding ~~ and ~=, the parameters

of the equivalent normal distribution once fitted at the linearization
point, we can solve for them as follows:

~, = P{l-l[FX(x*)])
s (530 )
M*)

(531 )

This process is illustrated in Fig. 5.2.

, Since }he linearization point xi*, changes with each iteration,

~X. and ~Z~ must be calculated for each iteration also. These values
are ;hen used in equations (5.15) through (5.17) as before. Note that
if the iteration is performed in reduced space, then distribution
transformation into reduced space has to be performed in each step.

f*(x)
A
\
ACTUAL DISTRIBUTION
w-
\

NORMAL DISTRIBUTION
\
\

X* x
LINEARIZATIONPOINT

Figure 5.2. Equivalent Normal Distribution at Linearization Point

135
 In our simple beam example discussed earlier, the load follows a Weibull
distribution while the strength was assumed to be normal. An equivalent normal
distribution can be determined for the load using the procedure described above.
The three limit state functions describing the failure of the beam are given by
equations (5.19) to (5.21). II one uses Hasofar/Lind method, then one must
determine the parameters of the equivalent normal distribution in each step of the
iteration procedure for the nonlinear limit state functions (equations 5.20 and 5.21).
The results for the ~ values including the distribution information
are [5.11]:

~~= 2.6690 ~ = 2.6690 ~3 = 2.6690

and the corresponding probability of failure is pf = @ (-~)= 0.003802.

The results indicate that the ~ values are invariant to the problem
formulation since ~1 = 13z= ~3. They also show that, in this case, inclusion of the
distribution information inmeased the ~ value from 2.5607 to 2.6690 and decreased
the corresponding probabli@ of ftilure from 0.005223 to 0.003802. The values of ~
and ~ deterrnined including the distribution information are more accurate. It
should be emphasized that inclusion of the distribution information does not
always yield larger safety index.

136
5.3 Correlated Random Variables

So far, the random variables %~ have been assumed to be

uncorrelated. If the variables are normal and correlated through a
correlation matrix [C], a transformation to a set of uncomelated
variables Yi is possible. The new uncorrelated set can be then used in
the procedure developed earlier for computing the safety index .
P

The set of uncorrelated variables

../ Y: can be determined from the
reduced variables ~-; using the orthogonal transformation

(5. 32)

where T
is an orthogonal transformation matrix and the superscript
U indicates the transpose. The transformation matrix ~ is such
that

Et c? J=A
[1 [1 (5. 33)

where [ C] is the covariance matrix of X and [ ~ ] is a diagonal matrix

of the eigen values of [ C]. The covariance matrix of X is to be related
to the covariances of the original variables ~ throug~

147
 11
P~~
12 P13
1
-----
-----
Pln
~2n
1
23
c+ ------- p3n
[11 31 32 1
I -------------- ------- 1
- - --- - 1
1 nl n2 n3
J (5. 34)

where
Cov IXiixjl
P. =
lj a u
x.
1 J (5. 35)

The safety index

P can be then calculated from equations (5. 15 to
5.17 ) for the new uncorrelated set Y or more directly from (see
reference [5.4]):

J-q-qz= (5. 3F)

where G* is a gradient vector evaluated at the linearization point

(most li~ely failure point), i.e.,

*+ ag ag ag
~z .* .*.
[ ax; ax: ax; 1
* (5. 37)

Notice that since ~ is orthogonal, (~-l =Tk)andfrom (5.32):

Et =~~
(538 )

and

(5. 39)

where

[1rx = Q
x
2

0 %=,
u
x
n

139
 Px
1

%= P
,X2
I

x
n
i 1

It can be shown (see reference [5.5] ) that the eigen values [~]
are also the variances of the variables YL .

For 1inear performance function g(x) represented by

g(x)=ao+ ai Xi
2 (5.40 )
i

equation (5. 36) for calculating ~ is reduced to:

*
where ~,, is given by (5..35).
d
The above procedure is valid for transforming a set of
correlated normal variables to a set of uncorrelated normal variables.
If the variables are non-normal, then as an approximation,
equivalent nokmal Yariables can be determined as described
previously under Inclusion of the distribution information. A more
exact procedure would be to use Rosenblatt transformation [5.6]
which requires information on the joint probability density function
of ~. The degree of approximation is illustrated through a numerical
example given in [5.5] and is usually small.

5.4 Rend of the =liahili~~ Ind= for Ei hteem Ships

First order reliability was used to calculate the safety indices

for eighteen existing ships. A linear performance function was used,
therefore the MVFOSM method yields the same results as the
Hasofer/Lind method. The margin M or performance function for
the primary strength is simply given by:

139
 M= 9-2

where S is the hull resistance given in terms of moment capacity and J

Z is the total applied bending moment which consists of stillwater
and wave moments. A worldwide mission profile was chosen for
sixteen of the ships as indicated in figure 5.3. The remaining two
ships have mission profiles shown in references [5.7, 5.8]. Three of
the eighteen ships are large tankers (190,000 dwt or larger), nine are
small to medium-size tankers (26,500 to 75,500 dwt), and the rest
are cargo and bulk carriers. Table 5.1 shows the general
characteristics of the eighteen ships. A strip theory program was
used in conjunction with Pierson-Moskowitz spectra [5.9] in order to
determine the root mean square and the mean values of the wave
bending moment in different sea conditions specified by their
significant wave heights. Figures 5.4, 5.5 and 5.6 show these results
for sixteen ships. The required results for the remaining two ships, a
Mariner and a tanker, were obtained from references [5.7] and [5.8],
respectively. Reference [5. 10] was used in order to determine the
frequency of occurrence of the different sea conditions. The mean
value of the wave bending moment was then obtained from the
mean values in the different sea conditions and the frequency of
occurrence of these sea conditions. A typical procedure is illustrated
in detail in references [5.7, 5.8]. The variance of the wave bending
moment was determined using the equation:

Variance = Mean square - square of mean value

In the computation, the following assumptions regarding the

sea description were made in order to reduce the computer cost:

(a) Pierson-Moskowitz spectra were used (fully developed

seas)
(b) Long-crested head seas were assumed.

140
 w-, .

70

V ---
$0
+--
40[- 63

30+ 30
I 17

i
20- 20

-.
10L 10

+ 91 f--:A+ ~<%--- . 0

!0

)X9
. 20
40
30
44 43
,--- 40

Figure 5.3. Assumed ships route (Figure without the indicated route
was obtained from reference [5. 10] )

141
 Table 5.1. General Characteristics of Eighteen Ships.
dwt
SHIP LBP (ft]B (ft) d (ft) Cb (approx. )
TankerNo.1 1076.IM 174.87 81.40 0.86
TankerNo.2 1069.25 163.25 Sa.o.i 0.R3
TankerNo.3 l$KJ.CUJ154.76 60.45 0.83
TankerNo.4 115.99 42.01O.fi
TankerNo.5 754:70 104.46U.40 0.82
TankerNo.6 754.70 105.654474 0.s09
TankerNo.7 7s.69 105.6544.74 O.MM
TankerNo.9 719.10 82.50 39.15 0.786 40; 970
Tuker No.10 620.81 35.72 0.784 31,500
TankerNo.11 .594.WI % E $.$ :.g 26, !ss0
TinkerNo.12* 775.W 105.50 75,500
ol&ep
700.65 9s.4340.70Oani 45,100
Cartiship No. 14 528.50 73.9929.SS 0.61.513,400
Cargo chip
No.15 520.00 75.
W 31,420.573 12,750
CargoshipNo.16 528.
W 76.0029.800.6W5 13,400
B::-cay
Uoo.
oo 106.00 44.550.840 74,200
Bu;ll
ce~
6ii6.2093.8042.630.793 ::,~:
T;nk;r
No.8 693.75 97.0039.170.77.5 , -
i Tanker and mariner ships from referenoeaf~,;~nd[s.~ $mWctivel y

1,000,000

900,000

800,000

i 700,000

4
600,000 TM4KER No. 1

12 KNOTS
500,000

400,000

300,000

6 KNOTS

200,000 i~

100,000

1 I I 1 I I
5 10 15 20 25 30 35 40 45 50
SIGNIFICANT WAVE HEIGHT [FT]

Figure 5.4. Rms of Wave Bending Moment for Three Large Tankers

142
 ~ 300,00( 6 KNOTS

WDIUM SIZE TANKERs

: 250,00( 6 TANKER NO. 4
? +- No. 5
o~ No. 6
200,000 12 KNOT% on No. 7
. No. B
v. No. 9
a. No. 10
150,000
16 KNOTS xx No. 11
+-

100,000

So,ooo

.[
> 10 15 20 25 30 35 40 45 50
significant WAVE HEIGHT [FTI

Figure 5.5. Rms of Wave Bending Moment for Eight Medium-Size

Tankers

I
. 300,000
OIL-OFX CARRIER MO. 13
c
CARGO SHIp No. 14
2- 15 KNOTS v CARI=O SHIP No. 15
250,000 4 BULK CARRIER No. 17
g :
+BULK CARRIER No. lE
x
+
u
z 200,000 b1
E 13.S KNOTS
z
:
150,000 20 KNOTS

:
+
m 100,000 4
/
i 1S KNOTS
i ; ~~
50,000

o 5 10 15 20 - 25 30 35 40 45 50
SIGNIFICANT wAw HEIGHT [ml

Figure 5.6. Rms of Wave Bending Moment for Cargo and Bulk
Carriers

143
 The strength coefficient of variation V$ was investigated next.
A detailed procedure for determining this coefficient is described
and applied to three ships (tanker, cargo, and frigate) in references
[5.7, 5.8]. The strength coefficients of variation of these three ships
were in the range of 7 to 11.3 percent. These figures include the
estimated subjective and the computed objective uncertainties. The
eighteen ships under consideration were assumed to have a strength
coefficient of variation = 13 percent. This is rather pessimistic and is
on the conservative side. The mean values of the strength, m~ were
considered to be equal to the section modulus of the ship multiplied
by the yield strength, taken = 30 k/sq in.

Obviously, the foregoing parameters for the eighteen ships

were determined in an approximate manner. More accurate
procedures can be used. However, the main objective was to provide
a preliminary investigation of the order of the magnitude of the
safety index ~ for as many ships as possible.

With these parameters determined for all the ships, the safety
index is computed from the equation:

=+
where 0 = ~5/mZ and V5 and V2 are C.O.V. of the strength and load,
respectively.

Figure 5.7 shows the computed values of the safety index of

?
these ships plotted versus the length between perpendiculars. These
results show a wide variation of the safety level, with ~ ranging
from about 4 to about 6.5. There seems to be a general tendency for
higher longitudinal strength safety, that is, higher , with increased
P
ship size.

The safety index P was also plotted versus several other

parameters. Figure 5.8 shows plotted versus the parameter
P

144
MS = pgL2Bd, which is proportional
to the static bending moment.
Only 15 ships are shown in this figure as the three large tankers 1, 2,
and 3 have very large MS values (470 x 106, 310 X 106, and 267 X
106, respectively) and would change the abscissa scale considerably.
They do however follow the same general trend of the data, which is
higher safety ~ for higher Ms. The physical significance of this figure
is that designing ships on the basis of the static bending moment Ms
would lead, in general, to higher safety for larger ships; that is, the
seems to overestimate the load on large ships.

t.
t

l?igure 5.7. Safety index ~ for Eighteen Ships

I I
10
I
20
1
10
I
40
I
50
I I I I
60 70 [0 so (x 10$ %

Figure 5,8. Safety Index ~ for Fifteen Ships

145
 Figure 5.9 shows the computed safety index ~ plotted versus
the parameter

(ml) Uy
F~ =
pg Lz Bd

Where (SM) is the elastic section modulus, cry is the yield

strength, p g is the water weight density and L, B and d are the
Iength, beam and draft, respectively. The parameter Fs is
proportional to a conventional factor of safety defined by dividing
the ship strength by the static bending moment. The figure shows
that some of the eighteen ships may have actually a lower safety ~
than the others even though Fs may indicate the opposite. This, and
the scatter in the data, suggest the inadequacy of the parameter Fs as
a measure of the real safety of ships. The same result was pointed
out in reference [5.8] for a more specific case.

: ~ %
6 1
f

)
t *
s _. :
s
: 1
.4
m
m :
n

I/
n
b

I I I I I I I 1 I I I
1.s 1.1 2.0 2.2 2.4 2.C 2.0 2.0 3.2 ).4 3.6 lx 10-% r,

Figure 5.9. Safety Index ~ for Eighteen Ships

In general, the wide range of the safety level of the eighteen

ships indicated by ~ as shown in figures 5.7, 5,8 and 5.9 could be due
to several reasons. The lack of allowance, or at least the lack of
uniform measure for the allowance of the uncertainties in the

146

656
bending moment and the strength in the traditional design procedures is a
possible reason. Another reason is the possibility ofinadequacy inthepredition
of the wave bending moment when using the traditional static balance
procedures. A third reason may be due to the slightly different niles of the
different classification societies used for dete rmining the section moduli of the
ships. In addition, the variation of the actual values of the section moduli of the
ships with respect to those specif5ed by the rules is also a contributing factor.
Finally, a more exact procedure for determining the parameters used in this
application may lead to a slightly different range of level of safety.

5.5 Shin Sa fetv~e x Based on Non-Linear Limit State Function;

The following analysis is presented to illustrate the use of Hasofar/Lind

procedure and the method of including the distribution information in calculating
the safety index for a ship. The following limit state fimction is considered

g(x) = (SM) fY- M~W- MW (5.41)

where
(SM) = minimum section modulus
fy = yield strength of the material
M Sw = stillwater bending moment
MW = wave bending moment

Notice that the product in the first term of (5.41) -makes the limit state
equation nonlinear. The limit state equation in the reduced space is

g(i) = (Vsm + ~srn (SM)) (Vfy+ afy f;) - (1.Lsw

+ ~s~ MSw)
(p*+awM~) = o (5.42)

where the prime superscript indicates a reduced variable and p and c indicate the
mean and standard deviation, respectively.

The derivatives required in evaluating the direction cosines (equation 5.17)

are:

147

Irl
 %rn (lJfY+ fY$) = %m fy (5.43)

Wy(m (5.44)

Ggw (5.45)

-% (5.46)

The direction cosines ~i can then be calculated from (5.17). Using these,
the coordinates of the most likely failure point are calculated from (5.16) and
substitut&in the limit state equation in the reduced space (equation 5.15) to yield
the following resulti

~%m - asrn ~*Srn~ ) ( ~fY- ~fy ~*fY~ ) - .

(vsw-aswa*swpb(~w-a~a*wp) = o (5.47)

Equation (5.47) is to be solved for ~. Notice that the U*i are to be evaluated at
the most likely failure point, i.e., fYand (SM) values at that point must be inserted
in equations (5.43) and (5.44), respective y, when evaluating ~*i given by equation
(5. 17). Since these coordinates are unknown apriori, the mean values of these
variables can be used as initial values in an iterative procedure. fir solving for
~ from (5.47), a new set of coordinates can be determined from (5.16) and used ta
determine a new set of ~*i. The procedure is repeated until conve~.nce is
achieved.

The above procedure is accurate if all the random variables in the limit
state function (equation 5.41) are normally distributed. However, as discussed
earlier, the wave bending moment MW follows a Weibull or an exponential
distribution. In this case, the mean and standard deviation of an equivalent
normal distribution must be determined in each iteration step according to
equations (5.30) and (5.31) and used in equation (5.47)prior to solving for ~. This

149
transformation will produce additional non-lineari@ in the limit state equation in
the reduced space.

For an exponential distribution with a parameter i, the equivalent

parameters (@ and pn) of the normal distribution for the wave bending moment
can be calculated from equations (5.30) and (5.31) which yield:

This procedure is further illustrated by the following numerical example.

Numerics 1 Examp~

consider a Tanker of the following characteristics:

Iagth = 763 ft Beam = 125 R Depth = 54.5 R

Draft = 41.33 R Blofik Coefficient = 0.805
Displacement = 90,650 L.ton DTW = 75,650 L.Inn

The values the means (@ and coefficients of variation (S) of the variable in
the Iimit state equation were determined as:

I&m = 150,441 inz -ft r)~~ = 0.09

Vfy = 18.16 tiid % = 0.10

I&w = 425,000 fhtin 8Sw = 0.35

l.1~= k = 150,500 ft-kan & = 1.0

For the first iteration in the procedure described above the mean values
were taken as the most likely failure point. The parameters for the equivalent
normal distribution for the wave bending moment can be calculated directly from
(5.48) and (5.49). The derivatives and timction cosines are then calculated from
equation (5.43 to 5.46) and (5.17), respectively. The resultig direclion cosines are
substituted in (5.47) yielding an equation in ~. The first iteration solution for ~ is
5.95. Using the determined value of ~ a new set of coordinates of the most likely
failure point is determined according to (5.16). The procedure is repeated and the
second iteration result~~a ~ value of 5.04.More iterationsshould be performed until
convergence is achieved. For comparison, if the information on the wave bending
moment distribution was not included, the resulting ~ values are 5.83 and 5.73 for
the first and second iterations, respectively.

150

Ibo
 lummcEs

5.1 Cornell, C.A., A Probability-based Structural Code, ACI Journal,

Dec. 1969.

5.2 Hasofer, A.M. and Lind, N. C., Exact and Invariant Second
Moment Format, Journal of the Engineering Mechanics
Division, ASCE, Vol. 100, No. EMI, Feb. 1974, pp. 111-121.

5.3 Rackwitz, R. and Fiessler, B., Structural Reliability under

Combined Random Load Sequences, Computers and Structures,
Vol. 9, 1978.

5.4 Shinozuka, M., Basic Analysis of Structural Safety, J. of

Structural Division, ASCE, Vol. 3, No. 109, March 1983.

5.5 Ang, A. H.-S. and Tang, W. H., Probability Concepts in

Engineering Planning and Design, Vol. II, John Wiley & Sons,
New York, 1984.

5.6 Rosenblatt, M., Remarks on a Multivariate Transformation,

Annals of Math. Stat., Vol. 23, No. 3, September 1952, pp. 470-
472.

5.7 Mansour, A.E., Probability Design Concepts in Ship. Structural

Safety and Reliability, Trans. SNAME, 1972.

5.8 Mansour, A.E. and Faulkner, D., On Applying the Statistical

Approach to Extreme Sea Loads and Ship Hull Strength, Trans.
RINA, Vol. 115, 1973.

5.9 Pierson, W. and Moskowitz, L., A Proposed Spectral Form for

Fully Developed Wind Seas Based on the Similarity Theory of
S.A. Kitaigorodskii, Journal of Geophysical Research, Vol. 69,
Dec. 1964.
5.10 Hogben, N. and Lumb, F., Ocean Wave Statistics, National l?hysical Laboratory,
Ministry of Technology, London 1967.

5.11 Skjolds, M., An Evaluation of Different Formulations of the Reliability

Index, Report for NA240B, Dept. of Naval Architecture and Offshore Engineering,
University of California, Berkeley, May 6, 1988.

152
 6. LEVEL I RELIABILITY 31WLYS1S :

Although Level 2 is easier to apply in practice, it is still of

limited use to practitioners. Normally a designer needs factors of
safety to apply in the design process such as those applied to the
yield strength of the material and to the loads. This need resulted in
the introduction of Level 1 reliability analysis. In this level partial
safety factors are determined based on Level 2 reliability analysis.
If these factors are used in a design, their cumulative effect is such
that the resulting design will have a certain reliability level (i.e., a
certain safety index). Thus, code developers and classification
societies may determine (and specify in their codes) these partial
safety factors that ensure that the resulting design will have a
specified reliability level. The method of determining these partial
safety factors for a given safety index is discussed next:

6.1 Derivation of Partial Safety Factors from Level 2 Method:

The partial safety factors (psf) or load and resistance factors

LRF are simply safety factors that are multiplied by the basic design
variables in order to assure a specified reliability level ~ . They are
usually applied to the mean values of the design variables, thus, w e
may write the limit state function as:

8 (Aipx ,A2PX , ..... Anpx ) =0

i 1 n (6.1)

where A; are psf and ~X: are the mean values of the variables. Since
equation (6. 1) represents the failure stmface, 4: ~z: must fall on the
,*
surface, preferably, the most probable failure point, i.e.,

$
x.
Ai=l
Px (6.2)
i
In the normalized variate space (see equation (5. 19 ) we may write

153

[0
Thus the original variate is
* *
x. = +axx?i=~ -af#ux
1 Pxi x.
i 1 i
= ~x(l-a~dvx) (6.3)
i i
Comparing equations (6.2) and (6.3) we conclude that

Ai=l-afpvx
i (6.4)

Evaluation of the psf Ai

requires evaluation of the direction cosines
x i at the design point, i.e., the most probable failure point t:. For
non-linear limit state functions, the determination of @~ requires an
iterative solution. The following simple procedure may b~ used [6.1]:
x; -
~xi
I. Assume x ~ and determine ~ ?~ = 1
i a
x:
2. Evaluate(~)=~d a:

* *
3. Determine i lx+axxbi~x-m~ fl~.
i i i i
4. Use, the new values of X; from step 3 again in 1 until
convergence is achieved.

5. Calculate the psf from Ais 1 - at P vxi for a given ~ .

Notice that if ~ is not prescribed but is to be determined, then

the above procedure can be -modified as follows. After Step 3, follow
the following steps:

4. Substitute x; determined in Step 3 into (~~, ,-- K; ) = o and solve

9
for .
P

154
 5* Using obtained in 4 above, reevaluate x ~ - mx
ip. *
P

6. Repeat Steps 2 to 5 until convergence is obtained.

6.1.1 Linear Performance Functions:

In this case the psf are such that

g ( AIPX , 9**.*
1
and

or isl
n
a. + ai Ai Px =0
1 i
i=l
Because of linearity the partial derivatives ~~~ are independent
of xi, that is i
ag
axti za.o
1
i

(since X.=a ~x+p

1
i i xi )

Therefore from (6.4), the psf are

,% a. c
1 x.
=1 ,- 1
i 1
pv
,
( aiax )2 T. i
[1 i 1 (6.5) .
i
6.1.2 Example

Consider the simple linear performance function

M= 8.(X1,X2) = s - z

where S represents strength and Z represents load.

In order to determine the psf for a given value of , we first

P
write the reduced variables as

155

lb~
 Z-pz
; z =
Z

and i3M
; az~ = - z

u a
therefore as = s z
i
and az = - 1
2 2 2
a + 0: T + a T
[ s 1 [ a z )

The psf are thus:

u
8

he=l-a~$~~=l-$v~ 1
2
a 2+a -r
[ s z 1

and u
z
AZ=l-UZ$VZ=l+$VZ 1
0 2+#;
s 1
T
[

Notice that At L ~ Wd AZ } ~ , as expected. The determination

of psfs for eighteen ships will be given in Chapter 10 of this
report.

6.2 Recently Developed Reliability-Based rorle.&

The procedure described above for the derivation of psf may

be used to develop safety factors for use in codes. This necessitates a
change in code format as well. Changing from a working stress
design code to a reliability-based code is not an easy task.
Complicating the procedure is the fact that there is no set method for
introducing reliability into a code. The implementation of reliability
theory in design codes changes from organization to organization.
Even when two organizations use the same reliability-based design
format, the details differ as it must for different types of structures.

156

/66
 It is the objective of this section to review and assess the
implementation methods and use of reliability analysis in certain
design codes. Specifically, work attributed to or sponsored by the
American Petroleum Institute, the National Bureau of Standards, and
Comite Euro-International du Beton are examined in some detail.
Codes of several other organizations are also discussed and
compared.

Proposed American Petroleum Institute Code Format

The work reviewed in this subsection is the proposed revision

to the API Recommended Practice for Planning, Designing and
Constructing Fixed Offshore Platforms (RP2A) which is issued by the
American Petroleum Institute (API) [6.2]. API is currently
sponsoring research aimed at changing the code format of RP2A from
working stress design (WSD) to load and resistance factor design
(LRFD), with the release of an LRFD-based RP2A for industry review
and comment envisioned in the near future. Much of the information
reviewed here that pertains to the proposed LRFD RP2A was
obtained from references [6.3, 6.4, 6.5, 6.6].

In the currently used working stress approach, the maximum

or yield stress is divided by a safety factor to obtain an allowable
stress. Designs are then limited so that the maximum calculated
stress under extreme operating loads does not exceed this. allowable
value. The basic safety checking format is of the form:

R
SF z D + L + W + other load effect= (6.6)

where

R = nominal component strength

SF = safety factor
D = nominal gravity load effects on components
L = nominal live load effects on components
w = nominal environmental load effects on components

157
 Presently, nominal loads are all combined with factors of 1.0,
and constant safety factors of 1.67 and 1.25 are used for operating
and extreme loading, respectively. Note that there is a probabilistic
statement implicit in the given safety factors, in that since extreme
events are by nature, rare, the associated safety factor can be
reduced.

Design based on the WSD format has provided structures with

high reliability without explicitly considering uncertainties and
probabilistic safety descriptions. The WSI) format, however, does not
provide for structures with uniform reliability. The problem with
WSD is that the one safety factor in equation (6.6) cannot possibly
account for uncertainties in all variables, including those arising from
the theories and analysis methods employed.

In the LRFD format, individual partial safety factors are

calibrated according to the different component strength and loading
uncertainties. The advantage of LRFD with its multiple factors is that
proper weight is given to the degree of accuracy with which the
various loads and resistances can be determined, resulting in a more
rational procedure and. a greater uniformity or reliability [6.7]. The
LRFD format recommended [6.3, 6.4, 6.6] for API RP2A has the form

Ri Ri>yDD+rLL+rw w+. ... (6.7)

where
nominal strength or resistance of component i
partial resistance factor for component i
nominal gravity or dead load effect
load factor- for dead load
L= nominal live load effect
load factor for live load
w= nominal environmental force effects with
prescribed return period (usually 100 years)
3 w= load factor for environmental load

158
 Each resistance factor 4 k; is calculated as the product of two
terms representing component strength uncertainty ( +: ) and system
consequence (+5 ), that is

*Ri = i = (6*8)

The component resistance factor #i takes into account variations

in material properties, strengths of fabricated components, and
errors in mathematical predictions of strengths (due to underlying
assumptions and approximations).

The system consequence factor reflects the relative

consequence on the entire structure of the failure of a component.
This, in turn, depends on whether the component was redundant or
not, brittle or ductile, main or secondary, etc. ~ In addition, the system
consequence factor covers any other social and economic impacts of
platform failure.

The load factors ~ are also calculated as the product of two

terms. These terms correspond to uncertainty in the load intensity
( y~ ) and uncertainty in the analysis required to calculate the load
effects (7A ). We then have,

~ Load effects = (J intensity) X ~ analysis) =zi~x~ (6.9)

,
In an actual design, the # - and ~-values would be tabulated,
and the design equation would be checked for all specified load
combinations. Actual derivations of load and resistance factors for
proposed use in API RP2A are explained in [6.8, 6.6].

Comite Euro-International Du Beton Code Format

The code discussed in this section is a joint effort of the Comite

Euro- International Du Beton (CEB), sometimes referred to as the
European Committee for Concrete, and the International Federation

1.!59
for Prestressing (FIP). The design code is entitled CEB-FIF Model
Code for Concrete Structures [6.9]. Information explaining the use of
probability in the CEB Model Code can be found in references [6.9,
6.10]. It should be mentioned that CE13 ~has been studying structural
reliability for many years and, for this reason, is considered a leader
in the field of such code development.

The CEB design checking equation has the general form

(6.10)

where ~~ is design resistance and s~ the design load effect.

The CEB code is a Level-I code, meaning appropriate levels of

structural reliability and provided by the specification of a number
of partial safety factors. The code development uses the Level-II
method whenever possible to assms appropriate values for the
safety factors used in the code. The code considers both ultimate and
serviceability limit states.

The format used for defining the design load effects, following
CEB notation, is

(6.11)

where

Sd = design (factored) load effects

s{... ) = refers to load effects due to all loads in brackets
(that is, it is not a numerical operator)
G= nominal dead load
P= a representative value of prestressing force
characteristic value of principal variable load
characteristic value of other less important variable
loads

160
xptfl~y= partial safety coefficients
poL = load combination factor

The CEB Model Code uses, in general, two types of partial safety
factors, x ~ and II , related to the strength and the loads [as in
equation (6. 11)], respectively.

yf is the partial safety factor which is multiplied by the

characteristic action (load) value, Fk, to obtain the design load effect,
that is
Design action = IJ t-K
- (6.12)

is a function of three factors: ~rl , ~~z , ~d 3 FJ .

accounts for variations in load magnitude from the

specified characteristic (that is, nominal) value. It is
analogous to the ~ i factor used by API.

reflects the reduced probability of combinations of loads

all acting at their respective characteristic values. r{ is
referred to as a load combination factor. [ l~Z is &tten
as @Ciin ~quation (6.1 l).]

accounts for the structural response to loads and the

possibility of redistribution of the load effects. ~~ reflects .
inaccuracies in predicting load effects, and is a ffnction of
the construction material, design and construction
process, and the limit state under consideration. In
equation (6.1 1), ~~ -values are written as ~j , ~f, and
1$ since the 7J -value differs for each load. Also in the
CEB nomenclature, F refers to a load in general while Q
refers to a variable load.

ym, the second type of partial safety factor, is used in the

structural analysis by dividing the characteristic strength of a section
( ~K) by Ym to obtain the design strength of the section, that is

161
 design strength = f k /x (6.13)
n

ym, accounts for any uncertainties in the predicted strength of

the materials used to build the structure. Specifically, it reflects any
variation in the strength of materials from the specified
characteristic value; any variation in the strength of the materials
from that predicted by control test specimens; possible flaws in the
structural material due to either the construction process or the
material itselfi dimensional inaccuracies of the material; and the
effect on the predicted structural resistance of inaccurate values of
material strengths. The ym partial safety factor is the CEB version of
a resistance factor such as~Ri used by the APL

There is an additional factor in the CEB code - the modifying

factor yn. This factor takes account of the inherent structural
behavior, that is, of parts of the structure which can fail without
warning, and the consequences corresponding to this failure. yn is
broken down into two factors:

yn 1 reflects the type of failure (ductile or brittle)

Yn2 accounts for the consequences of failure

yn is not used explicitly, but only modify ym or yf -values. In

[6.8, 6.6], actual values of some of these factors are presented.

National Bureau of Standards Code Format

The design code referred to in this section is American National

Standard A58, Building Code Requirements for Minimum Design
Loads in Buildings and Other Structures, published by the National
Bureau of Standards (NBS). Much of the information dealing with the
use of reliability in this Code was obtained from Development of a
Probability Based Load Criterion for American National Standards
A58[6.11].

162
 The format recommended for use in the A58 Standard is a
combination of the CEB-FIP format described previously, and the
load and resistance factor design format proposed by Ravindra and
Galambos [6.12]. Since the CEB format has already been described ,
The LRFD format of Ravindra and Galambos is explained together
with how the two formats are combined in an optimum way.
Information concerning the LRFD method of Ravindra and Galambos
can be found accompanying reference [6.12] in the September 1978
Journal of the Structural Division of the American Society of Civil
Engineers.

The LRFD criterion can be expressed as

M. Z YE(YDCD13m + YLCLL. + -yWcwwm + ...)

(6.14)

where
resistance factor
u.= nominal resistance
partial safety coefficients
deterministic influence coefficients
mean dead, live, wind loads, etc.

The terms representing the load effects (that is, the right-hand
side of equation (6.14) are defined as follows:

are deterministic influence coefficients that

transform the load intensities to load effects.
is partial safety factor representing the
uncertainties in structural analysis. It accounts for
approxima?,ions and assumptions in the underlying
theory and is somewhat analogous to the VA 10 ad
analysis factor used by API.
account for the degree of uncertainty inherent in
the determinations of the loads Dm, Lm and Wm.

163
 The major difference between the CEB and the Ravindra and
Galambos load representations is that the live load is a separate case
with its own load factor in equation (6.14), but is a multiple of the
maximum load ~0~ ~i~ in equation (6.11). NBS believes that the
computational simplification realized by expressing the arbitrary-
point-in-time load as per equation (6.11) will outweigh certain
advantages due to the increased accuracy of having a separate
loading case in equation (6.14).

The CEB format is not considered advantageous in other ways,

however. If the methodology of the CEB format was applied to a
situation combining dead, live, wind, and snow loads, a total of 32
loading combinations is possible [6.11]. On the other hand, the LRFD
method has only four combinations to be considered. Since it is
desired to explicitly state just a small number of fundamental load
combinations for simplification in design, the LRFD method is the
optimum choice in this regard.

The NBS format for load factors is therefore a combination af

the best features of two methods. However, the NBS format for load
factors follows LRFI) much more closely than it follows the cEB
format. The CEB method is used such as in the case of factored
arbitrary-point-in-time loads, mentioned earlier when comparing the
two methods. The design equations for load effects. take the LRFD
form, however, as it is simpler to use in the design process. Also,
from now on, the NBS format will be refereed to as LRFD.

In the case of resistance factors, # , the LRFD method is used

and not the CEB concept of using material partial safety factors, ~m .
This factor is closely analogous to the factor used by API. The
resistance factor, always less than unity, accounts for variability in
member strengths due to assumptions used in determining the
resistance equations, variability of material properties, variability of
dimensions, uncertainties in fabrication, and importance of the
component to the structure.

164

17+
 In summary, the format recommended by NBS is

(6.15)

NBS has tabulated many y-values for various materials and

loading combinations, and has outlined a procedure to determine
values consistent with an organizations objectives. These values of
y and the ~ determination procedure are discussed in references [6.8,
6.6].

Code Formats of Other Organizations

The American Institute of Steel Construction (AISC) released, in

September 1983, a proposed design code entitled, LRFD Specification
for Structural Steel Buildings, for the purpose of industry trial and
review. The code is based directly upon the LRFD method of
Ravindra and Galambos which represents a prototype for a new
generation of structural design codes. The implications of the
proposed change of the AISC code to the LRFD format will be felt by
other organizations which use this code in some way. The American
Petroleum Institute is one such organization as the current working
stress based API RP2A adopts several of its design provisions
through explicit reference to the AXSC Code.

Another organization which uses an LRFD procedure is the

American Concrete Institute (ACI). In fact, ACI introduced split load
factors to North American design codes back in 1963. The history of
the use of reliability in the AC1 code is presented by MacGregor
[6. 10]. In [6.8], the ACI method of deriving partial safety factors is
discussed.

The National Building Code of Canada [6.13] uses the following

split load factor format for load effects:

Load effects = S{T@ + ~(Y~L + 7,0W + y~~]~ (6. 16)

165

llj
where

s{... } = load effects due to all loads in the brackets (that is,
it is not a numerical operator)
TDfm? = load factors
D,L 9**. = are the loads (dead, live, . . .)
v= load combination probability factor equal to 1.0,
0.7, or 0.6 depending upon whether one, two, or
three loads are included within the brackets
m= 1.25 if D acts in the same way as the loads in the
brackets and is 0.85 if D acts in the opposite way.

The load factors account for variations in the load effects due to
model errors and uncertainties in the structural analysis. The v term
reflects the reduced probability of maximum dead, live, wind, etc.,
loads acting simultaneously. Note that if both live and wind loads
were present, equation (6. 16) would design using the entire wind
effect (depending, of course, upon the chosen load factor). The LRFD
procedure and the CEB method are bcth considered more flexible
than the format of equation (6.16).

Another organization which uses a split load and resistance

factor format in a design code is ~et norske Veritas (DnV) in their
Rules for the Design, Construction, and Inspection of Offshore
Structures [6. 14], although the partial factors are not reliability
based. The format is somewhat similar to other European codes such
as CEB. The general format for ultimate limit state design is

(6.17)

where

Rk . characteristic resistance (strength)

166
 Ym = material factor
Rk/ym = design resistance
K= factor depending on type of resistance
S[z F:Y~i} = design loading effect
Fi = characteristic load
Yfi = load factor

The material presented in the foregoing subsections gives a

description of the major reliability-based code formats. There are, of
course, other formats, but these are generally variations of the
formats given herein. The format used more widely than any other,
at least in North America, is the load and resistance factor design
method. Whether it be the procedure proposed by Ravindra and
Galambos, or a variation thereof, LRFD seems to be a good practical
way of incorporating reliability into a design code. Most importantly,
an LRFD-based code is the simplest to use in practice, and this may
be a major consideration.

Deciding upon a design code format which allows for the

implementation of reliability methods is a complicated process.
However, once an organization has chosen a reliability-based code
format, the work is by no means over. A load and resistance factor
design code format may look quite impressive, but is of very little
use without the corresponding load and resistance factors.
,..

Generally, the first step in deriving partial load and resistance

factors for use in the LRFD format is to calibrate these factors based
on the reliability level inherent in the current design criteria. From
this inherent reliability, a corresponding target reliability level ~ is
established. Using this target value, load and resistance partial
safety factors are determined for the new format such that they
minimize the deviation of the calculated reliability from the target
level over the range of design applications. Although the target
reliability level cannot be reached for all design conditions, it should
be achieved on the average. Reference [6.8] gives more detailed

167

177
information on this procedure as well as some typical values of the
1oad and resistance factors corresponding to code formats developed
by the various organizations discussed previously. Examples and
additional information can be found in [6,8, 6.6, 6.3, 6.9, 6.11, 6.15].

168

11?
6.1 Ang, A. H.-S. and Tang, W. H., Probability Concepts in
Engineering Planning and Design, Vol. II, John Wiley & Sons,
New York, 1984.

6.2 API Recommended Practice for Planning, Designing and

Constructing Fixed Offshore Platforms, API RP2A, American
Petroleum Institute, 12th cd., Dallas, Texas, 1981.

6.3 Moses, F., Guidelines for Calibrating API RP2A for Reliability-
Based Design, API Prac. Project 80-22, American Petroleum
Institute, Oct. 1981.

6.4 Moses, F. and Russel, L., Applicability of Reliability Analysis in

Offshore Design Practice, API Prac. Project 79-22, American
Petroleum Institute, June 1980.

6.5 Anzari, T. and Bryant, L.M., Comparison of a Limit State Design

Code with API RP2A, Proceedings, 14th Annual Offs here
Technology Conference, Reprint OTC 4191, Houston, TX, May
1982, pp. 291-306.

6.6 Moses, F., Utilizing a Reliability-based API RP2A Format, Final

Report, API Prac. Project 82-22, American Petroleum Institute,
NOV. 1983.

6.7 Galambos, T.V. and Ravindra, M. K., Tentative Load and

Resistance Factor Design Criteria for Steel Beam Columns,
Research Report No. 32, Civil Engineering Department,
Washington University; St. Louis, MO, Feb. 1976.

6.8 Mansour, A.E., Harding, S. and Ziegelman, C., Implementation of

Reliability Methods to Marine Structures, First Annual Report
ABS, University of California, Berkeley, Jan. 1983.

169
6.9 Common Unified Rules for Different Types of Construction and
Material, Bulletin DInformation No. 124E, Comite Euro-
International du Beton, Paris, April 1978.

6.10 First, C)rder Reliability Concepts for Design Codes, Bulletin

DInformation No. 112, Comite Euro-International du Beton,
Munich, July 1976.

6.11 Ellingwood, B., Galambos, T., MacGregor, J. and Cornell, C.,

Development of a Probability Based Load Criterion for
American National Standard A58, NBS Special Publication SP-
577, National Bureau of Standards, Washington, DC, June 1980.

6.12 Ravindra, M. and Galambos, T., Load and Resistance Factor

Design for Steel, Journal of the Shuctural Division, ASCE, Vol.
104, No. ST9, Sept. 1978, pp. 1355-1370.

6.13 National Building Code of Canada, National Research Council of

Canada, Ottawa, 1977.

6.14 Rules for the Design, Construction and Inspection of Offshore

Structures, Det norske Veritas, OS1O, 1977 (and Appendices).

6.15 Yura, J., Galambos, T. and Ravindra, M., The Bending Resistance
of Steel Beams, Journal of the Structural Division, ASCE, Vol.
104, No. ST9, Sept. 1978, pp. 1355-1370.

169a
(THIS PAGE INTENTIONALLY LEFT BLANK)

.x
7, SIMULATION AND THE MONTE CARLO METHOD:

7.1 General concept

In general, simulation is a technique for conducting experiments in a

laboratory or on a digital computer in order to model the behavior of a system.
Usually simulation models result in simulated data that must be treated
statistically in order to predict the future behavior of the system. In this broad
sense, simulation has been used as a predictive tool for economic systems,
business environment, war games and management games.

The name Monte Carlo method was introduced in 1944 by von Newmann
and Ulam as a code name for their secret work on neutron diffusion problems at
the Los Alamos Laboratory [7.11. The name was chosen apparently because of the
association of the town Monte Carlo with roulette which is one of the simplest
tools that can be used for generating random numbers.

.Monte Carlo simulation is usually used for problems involving random

variables of known or assumed probability distributions [7.21. Using statistical
sampling techniques, a set of values of the random variables are generated in
accordance with the corresponding probability distributions. These values are
treated similar to a sample of experimental obsemations and are used to obtain a
sample solution. By repeating the process and generating several sets of
sample data, many sample solutions can be determined. Statistical analysis of
the sample solutions is then performed.

The Monte Carlo method thus consists of the three basic steps:

a. Simulation of the random variables and generation of several sample data

using statistical sampling techniques

b. Solutions using the sampled data

c. Statistical analysis of the results

170
 Since the results from the Monte Carlo technique depend on the number of
samples used, they are not exact and are subject to sampling errors. Generally
the accuracy increases as the sample size increases.

Sampling from a particular probability distribution involves the use of

random numbers as will be discussed later. Random numbers are essentially
random variables uniformly distributed over the unit intenal [0,1]. Many codes
are available for computers for generating sequence of pseudo random digits
where each digit occurs with approximately equal probability. The generation of
such random numbers plays a central role in the generation of a set of values (or
realizations) of a random variable that has a probability distribution other than
the uniform probability law.

The Monte Carlo method is considered now as one of the most powerful
techniques for analyzing complex problems. Since its chief constraint is
computer capability, it is expected to become even more commonly used in the
future as computer capabilities increase and become less expensive to use.

~ .2 fM r 1 li ili nli

As was discussed in Chapters 1,4 and 5, the reliability of a structure can be

~characterized by a.lirnit state function g@ = ~x1,x2 . . . xn), where xi are random
.
variables representing the basic design variables. The inequality g(~ ~ O
corresponds to failure, while g(~ s O represents the safe region. In the Monte
Carlo approach a random sample values xi for the basic design variables is -
generated numerically according to their probability distributions using a
random number generator (see the following section). The generated sample
values are then substituted in the limit state function whose value is then
computed to see if it is negative or positive, i.e., failure or no failure. Repeating
this process many times, it is possible to simulate the probability distribution of
g(~. TMs will require a very large number of samples. The probability of failure
can then be estimated from either of the following methods:

a. The probability of failure is given by

171
 (7.1)

where N is the tatal number of trials or simulations and n is the number of trials
in which g(~ ~ O.

The ratio nLN is usually very small and the estimated probability of failure
is subjected to considerable uncertainty. In particular the variance of n/N
. depends on the total number of trials N, decreasing as N increases. That is, the
uncertainty in estimating pf decreases as N increases. Statistical rules can be
used to establish the necessary number of trials which depends on the magnitude
of p~ Many variance reduction techniques have been developed to decrease the
variance of n/N with smaller number of trials than would have been necessary
otherwise.

b. In the second method, the probability of failure is estimated by, first fitting
an appropriate probability distribution for g(~ using the trial values described
earlier [7.3]. The moment or any other established statistical method may be used
in the fitting process. Elderton and Johnson [7.4] suggested some distributions
that are suitable for fitting the g(~ data. The probability of failure is then
determined from

Pf = ~f~ (m) dm (7.2)

where M = g(~) is a random variable representing the margin and fM(m) is its
probability density function as estimated from the fitting process.

7.3 Ge neration o f Rando m Numbers F or a IUmdom Van able With a Presc rib~
(-lontinuou~ Prob abilitv Distribution:

As mentioned earlier the process of generating random numbers with a

specified probability distribution may be accomplished by first generating
uniformly distributed random number between O and 1. Through appropriate
transformation, one may then obtain a corresponding random number with a

172

fw
specified probability distribution. Therefore, in this section we will first discuss
how to generate uniformly distributed random numbers then how to obtain the
corresponding random numbers with a specified probability distribution.

7.3,1 Random Numbers With Uniform Distributions~

Special devices may be used to generate uniformly distributed random

numbers within a given range. For example, by equally subdividing the
circumference of a wheel into a number of intewals equal to the given range and
spinning the wheel, the desired uniformly distributed random number can be
generated. Uniformly distributed random numbers are also tabulated and are
available in the literature for pencil-and-paper Monte Carlo simulations.

In computer simulation, methods for generating uniformly distributed

random numbers are generally based on recursive calculations which, because of
cyclic effects, do not generate truly random numbers. The generated set
eventually repeats itself after a very long cycle and, therefore, referred to as
pseudo-random or quasi-random. An example of a recursive calculation of the
residues of modulus m* that produce such a set of pseudo-random numbers is

Xi = a xi.1 -t b (mod m) (7.3)

where a, b and m are nonnegative integers and the quotients xi/m constitute the
sequence of pseudo-random numbers [7.1, 7.21. Such a sequence
repeats itself
after almost m steps, i.e., cyclic. For this reason m must be set very-large e.g; 108
or larger.

7.3,2 Random Numbers with Prescribed Distribution

Based on a generated set of uniformly distributed random numbers between

O and 1, one may obtain the corresponding random numbers with a specified
probability distribution. This can be done using a method known as the inverse-
function method. The method is suitable if the inverse of the prescribed
cumulative distribution function C.D. F. of the random variable can be
expressed analytically. The method is illustrated as follows.

173
 Suppose that a set of random numbers are to be generated fora random
variable Y which follows a prescribed distribution with C.D.F. I?y(y). The value of
Y at F~y) = x is
Y= FY-l(x) (7.4)

where FY-l(x) is the inverse of the C.D.F. at x. If X is a uniformly distributed

random variable r.v. between O and 1, then

Fx(x) = X O<x<l

Thus if x is an outcome of the r.v, X, the corresponding value of Y obtained

from (7.4) will satisfy the following equations.

P[Ygy] = PIFy-l(X)syl =P[X~Ffiy)] = l?x[F#y)] = F~y)

This means that if (xl, X2 . . . xn) is a set of numbers of the r.v. X, the
corresponding number obtained from equation (7.4), i.e.,

Yi = FY-l(xi) i=l,2, . ..n {7.5)

will have the C.D.F. Fy(y) as required. As an example consider the r.v. Y to have
a Weibull distribution with C.D.F. given by

The inverse function is

Y= FY-l(x) = k[-ln(l-x)]~~= k[-lnx]~! (7.7)

since x and (1-x) have identical

distributions, i.e., uniform distribution. Thus one
can generate the random numbers yi, i = 1, 2, . . . n corresponding to uniformly
distributed random numbers xi according to (7.7) from:

Yi = k[-lnxi]~? (7.8)

174
 The Weibull distribution can be reduced to the Rayleigh and the exponential
distributions as discussed earlier. If the Weibull parameters k and ~ are equal to
~E and 2, respectively, the Weibull distribution reduces to the Rayleign
distribution. If k and 1 are equal to L and 1, it reduces to the exponential
distribution. Thus, substitution for k and! in equation (7.8) w-ill lead to a set of
random numbers for these two special distributions as well.

7,4 Sanm le Size and Variance Reduct i on Tec hniaue~

As mentioned earlier, the simulated data according to Monti Carlo method

should be treated as a sample of experimental observation, and therefore, is
subjected to sampling error. If the probability of structural failure is to be
computed from the simulated data the underling error becomes an important
consideration since the sought probability of failure is usually small. Shooman
[7.5] developed the following expression for estimating the error in the estimated
probability of failure:

l-pfl12
error = (7.9)
z-
()

where N is the total number of simulations (sample size) and pf is the probability
of failure. There is a 95% chance that the actual error in the estimated probability
is less than that given by equation (7.9). It is seen that the error is dependent on
the number of simulations N and the probability of failure p~ it decreases by
increasing N or pf. Therefore, if the estimated probability pf is small which is
usually the case, N should be large enough to decrease the error.

There are techniques, however, which may reduce the error (or variance)
without increasing the sample size. These techniques are known as variance
reduction techniques, and the one that is used often in stmctural failure problems
is called Antithetic Variates.

7.4.1 Antithetic Variates

Let Y1 and Y2 be two unbiased estimates of Y as determined from two

separate sets of samples or simulation cycles. The average of these two unbiased

175
estimations Ya = l/2 (Yl + Y2) is also an unbiased estimator since its expected
value E [Ya] is equal to Y. The vafiance ~~a2 of the new estimator Ya is
determined from the individual variances OYIZand OY2Zas:

ayaz = U4 [ OY12+ @ + 2 Covml, Y2) 1 (7.10)

If Y1 and Y2 are negatively correlated, i.e., the COV. (Yl, Y2) <0, it is seen
from equation (7.10) that the third term becomes negative and

Oyaz < l/4 (CY12+ UY22) (7.11)

That is, the accuracy of the estimator Ya can be improved (or its variance
can be reduced) if Y1 and Y2 are negatively correlated estimators. The antithetic
variates method is thus a procedure that ensures a negative correlation between
Y 1 and Y2. This can be accomplished in structural reliability problems as
follows.

If X is a random vatiable uniformly distributed between Oand 1, then 1-X is

also a uniformly distributed random variable between O and 1 and the two random
variables X and 1-X are negatively correlated. Each of these random variabl~s can
be then used to generate the basic random variables Yi which have prescribed
probability distributions as described earlier. This results in a pair of negatively
correlated basic random variables. The procedure is repeated for all the random
variables Yi in the limit state equation. The limits stite equation is then solved for
each negatively correlated set of random vafiables separately and the results are
averaged to estimate the population mean. Note that the emor (or variance) of the
result is reduced because of the negative correlation between the generated
variables according to equation (7.11).

7.5
~J

71mri
~ of Anal i 1 n im 1 Ev 1 ion f a
Function
176

188
 The purpose of this example is to compare numerically simulated results
obtained using a standard random number generator with exact analytical
values. For this purpose, the random process x(i)(t) was examined in ref. [7.6]

x(i)(t) = Asin(at +6(i)) (7.12)

where A and o are fixed (deternxinistic) constants and e(i) is a random phase
angle with uniform distribution shown in Figure 7.1.

Analytical Results

The analytical result for the first order and joint probability density
functions (jpdfl of x given that e is uniformly distributed can be derived by
standard statistical methods. These are given by
f(x) = l/ CHJQ2 - x~l -A<x<fq

= 0 otherwise
and

f(xl ,x2) = l/[~~J92 ~lz]

A <(X1,X2)< A
C S(X2 xlcasa~ + dA2 - x12 s.inur)
- 5(X!? - Xl C05i.
U~ - J~2 - X12 ~~jq~~)]
(7.13)
=
Otherwise

where ~ = ~ - tl. -0
,

The first order pdf of x is shown is figure 7.2. In general, the jpdf is ~ -
difficult to represent graphically. It is three-dimensional with spikes when the
argument of either of the two delta functions is equal to zero. The occurrence of
these spikes will depend on the values of A,o and z as well as the c,urrent values of
x 1 and X2. The factor in front of the delti functions modifies their sum so that the
total area underneath the jpdf will equal one.

Numerical Results

First Order PM For given values of A, a and ~, a data file of N random

phase angles and the corresponding values of the x was created. The values of x
were generated by simply substituting the random values of e into equation (7.12).

177

/$?
BASIC language was used to create the data file, because it contains a random
number generator. A separate routine was written in FORTRAN to compute the
pdfofeither the random phase angle ortherandomxs. This routine divided the
range from O to 2n for 6, or from -A to A for x, into n intervals of size A6 or Ax.
Then the number of 9s or xs in each intenal was counted. The value of the pdf
at the center of each intenal is given by,

f(.) = ( #of occurrences per interval)/ [ ( Ae or Ax ) * N ]

where N is the total number of samples (simulations).

The probability density function of theta compares well with the expected
uniform distribution, see figure 7.3. The pdf of x was computed for the case of

A= 1 magnitude unit
a = 1 rad / (time unit)
t = 1 time unit
N= 100000 samples

The numerically computed points compare extremely well with the analytical
curve, see figure 7.4. There is, however, disagreement between the computed
talues and the analytical curve at the singularities, x = Q.

Second Order PDF: The numerical procedure used to calculate the jpdf was
very similar to that for the first order pdf.

Some easily visualized cases of the jpdf were investigated. Consider the
case when tl = t2 so that w = O. Since sin(0) = O and COS(0)= 1, the argument of
both delta functions in the analytically derived result (equation 7.13) simplify to
(Xl - x2). This implies that the jpdf will consist of spikes along the line, xl = X2.
The numerical results for this case with

A= 1 magnitude unit
N= 100000 samples
Axl, Ax2 = 0.05 magnitude units

178

(fo
is given in figure 7.5.

Another interesting case is when a = 1 rad / (time unit), tl = 1 time unit and
~ = 2.57 time units, then a = tiz. Since sin(n/2) = 1 and cos(ti2) = O, the argument
of both delta functions will equal zero when X12 + X12 = A2. This means that the
jpdf will be a series of spikes along a circle centered at (xl, x2) = (O, O)and of radius
A. The numerically computed jpdf evaluated under this condition is shown in
figure 7.6 and it indeed appears as expected.

Effect of Number of Samples and Size of Increment: The analytical

derivation of the probability density function assumes an infinite number of
samples and infinite resolution of x. In a numerical computation, however, both
N and AX are finite values. On the average, the number of samples per interval
will equal the total number of samples divided by the numbers of increments, i.e.,

# samples per interval = N / [ 2MAx ].

The trend shown in figure 7.7, as well as, in figure 7.8, shows that as N increases
for a constant Ax, the numerical values converge to the analytical result for the
pdf. In addition, comparing figure 7.7 to figure 7.8 for a constant N, shows that
decreasing the size of the interval Ax appears to increase the spread of the
fiumerical data. Therefore,. in order to get an accurate numerical representation
of the probability density function, not only the total number of samples and the
size of the increments are important, but mainly their relationship in
determining the number of samples per internal is important.

179
 Probability Density Function of Theti

l/2rr I
--- . _..

j
1
(

+ --
0 Theta

Figure 7.1. Analytically Derived PDF of Theta

Probability Density Function of X

I I

i
i

!
!
I i
(
I ) I
1
i I
I 1
I
; I
I I
I I
I I
I I., I
I
+-- L------- - --
1
-[
-~
-A A
~fi)
= .A(sinrt +- Gti))
Figure 7.2. Analytically Derived PDF of X

180
 Probability Density Function of Theti
100000 Samples. A9 = 0.01

0.2

1/2n

0.1

3.0
0 in
Tineta

Probability Density Function of Theta

100000 Samples. A6 = 0.1

0.2

1. ... ., . .
l/21r
I * -- .-. . .
- -*--

0.1 +
I

I
!

0.0 t . 4
0 Theta 2%

Figure 7.3. Numerically Computed PDFs of Theta

A9 = 0.01 and AO = 0.1

181
 Probability Density Function of x
(A, a, t) = (1,1, 1)
100000 %Ples Ax = 0.01

.

1.5

1.0 1

0.5

)
0.0 1 I [

-1.5 1.0
I 1 1
-0.5 0.0 1.0 1.s
Xfi) = Asjn (~ + ~;i))

Probability Density Function of x

(A, a, t) = (1,1,1)
100000 Samples Ax = 0.1

1.5

1.0

0.s

I
I
0.0 1 I 1~1
. I -1.0 -0.5 0.0 . 1,5
xfi)= ,%in(vt + ~$

Figure 7.4. Numerically Computed PDFs of x

Ax = 0.01 and Ax = 0.1

182

w
 2.5
r
2.0
t
L

1.5
Ill

&lgure 7.5. Joint Probability Density Function

(A 0, tl, tz) = (1,1,1,1)

1.0

.1*OL

Figure 7.6. Joint Probability Density Function

@~, tl, b) = (1,1,1, 2.57)

183
 1
. Probability
1000
4%%- %?-
Ocn~iiy Funct~on
0.0!
of X
20
Probability

10000 $:;/.:
Density

W)-
Function

0.0,
of X

IS

Lo -
I .
.
.
1.3

1.0

os- . 03

.*. . -..-**.

H t --- J -*****-*~
.. ---- ***-**.

04 r T 4
0.0d r I 1

-15 -1,0 -0.s 0.0 03 1.0 13 -{3 -1.0 -0.s O.kl 1.IY 1.s
f

Xt) = ,%in(Ft + @ X~ - Asin(rt + &0)

N = 1000 N= 10000

Probability Density Function of X

2.0
50000 $%)- W- 0.0! . .
1. .

1.5-
. . .
.
.


1.0-

0.0
-13
1
-1.0 -0.5
I .
1
1.0 1.
0.5-

O.b1
-ls
L..J
-1.0

-6.s
*.

I
0.0
.

O.k

1.0
1
1.5
Xo) m A&t + ;0? X~ = hin(~t + Qq

N= 50000 N= 100000

Figure 7.7. Ax = 0.01

[q6
 Probability Density Function of X Probability Dermity Function of X
2.0
1,00 !!%;)*- vi)- ~.,

1
10000 !%;/.; ii)- 0.,

I
1.5i I
. ., .
I

t.o 7

0.5-
.
. j ,y_J
1

0.0 1 t 1 f

-1.s -1.0 -0.s 0.0 lh 1.3 -1.5 -1.0 -0.5 1.0 1.5
X(O _ ~i~(~ +0~0~ Xfi)s Aj~(a~ +~~i~

N = 1000 N= 10000

Probobili& Density Function of X Probability Demity Function of X

Lo Somd4%$: w)- 0.1 2.0
1 1 1000.0
&:&- ii)-
0.,

1.5-

.

1.0-

0.s -

0.0
1 I ( I I
1 0.0
-1.5 1.: r
-1.0 -0.5 0.0 1.0
-13 -1.0 -d.s 0.0 ok 1
1.0 1,1
X@)- fijn(~ +o~(i)) Xm . Asin@ + G(IO

N = 50000 N= 100000

Figure 7.8. Ax = 0.1

185

177
?,5,2 A~nlication of Monte Ca rlo Method to Reliability of Shi~ Structures
(excerpted from White and Ayyub [7.7]):

In order to compare results of the Monte Carlo simulation method with the
other reliability methods discussed earlier in Chapters 4 and 5 an example
problem is solved using each method. The problem chosen is to determine
probability of ductile yielding of a vessels deck under extreme bending moments.
Any of the other possible modes of failure could have been chosen, for example,
plastic collapse or buckling, but the availability of data on this problem facilitated
comparison of methods. The vessel chosen for the analysis is a naval frigate, the
same one used by Mansour and Faulkner [7,8]. The principal dimensions are
given in Table 7.1 and the midship section is shown in Figure 7.9.

The problem is essentially a simple beam in bending and can be written as:

Mu = CY (7.14)

where Mu is the ultimate bending moment; C is the section modulus of the vessel;
Y is the tensile yields stress of the vessel material.

In order to see the effect of different, but mechanically equivalent

formulations on each method two limit-state equations will be used. The first
limit-state equation has a simple linear form:

Z= R-Q (7.15)

where R is the resistance, given in tons/in2 and is equal to Yin Equation (7.14); Q
is the total load in tondinz .

Next a more complicated non-linear form is used. This form separates the
wave and still water bending moments, Mw and M. respectively; and Z is .
expressed in units of bending moment:

z =YC-MO-MW (7.16)

186

Ftl
The basic variables for each form are shown in Table 7.2 along with their
respective statistical properties.

First Order Reliability Method (Level 2): The method described in Chapter 5
was applied in [7.7] to the linear and non-linear limit state functions given by
equations (7. 15) and (7. 16), respectively. In both cases the distribution information
was included in the analysis. The results for the safety index ~ and the
corresponding probability of failure p~ are given in column 2 of Table 7.3.

Direct Integration Method (Level 3): The method described in Chapter 4

was applied in reference [7.8] using the linear limit state equation only (equation
7.15) and assuming the stillwater bending moment to be deterministic. In
reference [7.8], the probability of failure was computed for ship operation period of
twenty one years (n=3). To be consistent with the results given in this example,
the value given in [7.81 must be divided by 3 and the corresponding resulting value
for pfis 1.3x 10-6 which is shown in Table 7.3.

Monte Carlo Simulation Method: In reference [7.7] the same problem

(equations 7.15 and 7.16) was solved using Monte Carlo simulation technique
described in this chapter. TJse was made of Antithetic Variate reduction method
and conditional expectation in order to reduce the number of simulation cycles.
The primary steps involved in the solution according ta reference [7.7] are:

step 1. Identify the basic variable with the most variability in the limit
state equation.
step 2. Condition the variable in Step 1 with respect to all the remaining
variables in the limit state equation.
step 3. Generate a uniformly distributed random deviate for each of the
conditional variables.
Step 4. Generate a second uniformly distributed random deviate which is
negatively comelated to the one from Step 3.
Step 5. Using the inverse transform method produce a random variable
for each deviate from Step 4.
step 6. Calculate the probability of failure using the probabilistic
characteristics of the variable identified in Step 1 for each set of
random variables.

187
 step 7. Find the average probability of failure for the two pfs in Step 6.
Step 8. Repeat Steps 3 ta 7 N times.
step 9. Calculate the statistics of the N number of probabilities of failure
thus generated.

The results are shown in Table 7.3 for 2000 simulations cycles. Figures 7.10
and 7.11 from reference [7.7] show the simulation scheme converges on a solution
with increasing N.

188

200
 ..

.I
J*

.-

[idship Section of Fngati of 360 R LBP [7.8]

189
 !.&,&,.,,,=
0.

0.s
0 12m 1-

Figure 7.10. Probability of Failure vs. Number of Simulations

Using Variance Reduction Techniques

190

202
0.1

0.

0.

0
0
~ w Sxunv%w Cvcum

Figure 7.11. COV of Probability of Failure vs. Number of

Simulations Using Variance Reduction Techniques

191
 Table 7.1. Vessel Characteristics

Length Between Perpendiculars 360.0 R (11O.OOm)

Beam (molded) 41.0 R (12.50 m)
Depth 28.9 ft (8.78 m)
Draft 12.0 ft (3.66 m)
Displacement 2800.0 tm.s (2845.00 tmnes)
Section Modulus (at deck) 5700,0 inz ft (1.12 ma)

Table 7.2. Probabilistic Characteristics of Basic Variables

Basic Variable Mean Cov Di4tibutiou

Linear Formulation
R 22.20 tonahn2 .0710 Normal
Q 2.70 tanshz .5390 Weibul
Non-Linear Formulation
Y ~ 22.20 tonsAn2 ,0610 Normal
c 5700.00 in~ft .0379 Log-Normal
M. 7080.00 R-tins --- Deterministic
Mw 8290.00 fbtins Loom Weibul (k=l)

192

20$
 Table 7.3. Example Problem Resul&

First Order Direct Integration Monte Carlo

Method Method Method

I
Linear Limit p = 4.75
Stab Equation (7.15) pf = 0.97x 10+ pf = 1.33 x 10+ M = 0.98X 10~
(C*O.V. = 0.0192)

Non-Linear Limit p = 4.75

State Equation (7.16) pf = 0,976X 10~ ..- M = 0.98X 10~
(C.o.v. = 0.0174)
 ENCES

[7.1] Elishakoff, L, J?robabilistic Methods in the Theom of St~c ture~, John Wiley
and Sons, 1983.
[7.2] Ang, A. H.-S. and Tang, W. H., Probability Concepts in Engineering
Planning and Design, Risk and Reliability, John Wiley and Sons, New
York, 1984.
[7.3] Theft-Christensen, P. and Baker, M., Structural Reliability Theory and Its
Application, Springer-Verlag, Berlin, 1982.
[7.4] Elderton, W. P. and Johnson, N. L., Systems of Frequency Curves,
Cambridge University Press, 1969.
[7.5] Shooman, M. L., Probabilistic Reliability: An Engineering Approach,
McGraw-Hill Book Co., New York, 1968.
[7.6] Ochs, J., Comparison of Analytical and Numerical Evaluation of the
Random Function x(i) = A sin(uz + e), Report for NA 240B, University of
California, Berkeley, May 1988.
[7.7] White, G. J. and Ayyub, B. M., Reliability Methods for Ship Structures,*
Naval Engineers Journal, Vol. 91, No. 4, May 1985.
[7.8] Mansour, A. and Faulkner, D., on Applying the Statistical Approach b
Extreme Sea Loads and Ship Hull Strength, Trans., Royal Institute of
Naval Architects, Vol. 115, 1973.

194
 8. SYSTEM RELLM31LI~

8.1. Introduction

The reliability analysis discussed in Chapter 4, 5 and 6 has been mainly

concerned with a single failure mode (or a limit state) defied by a single limit
state equation. Marine structures, however, involve several modes of failure, i.e.,
there is a possibility that a structure may fail in one or more of several possible
failure scenarios, The subject of system reliability deals specifically with the
methods of combining the probabilities of failure associated with these modes in
order ta determine the total reliability of the stmcture as a system.

Two main sources of system effects are identified. The first is due to
possible multiplicity of failure modes of a component or a structural member. For
example, a beam under bending and axial loads may fail in buckling, flexure or
shear. Each one of theses modes can be defined by one limit state equation. Even
though in this case, we are dealing with a single member (beam), system
reliability methods must be used in order to combine the possible failure modes
and to obtain an assessment of the total risk of failure of the beam. The probability
of failure of one mode may be larger than the others, but the fact that there is a
possibility that the others may occur indicates that they must be included and
combined ta obtain-the tot@ probability of failure of the beam.

Another example of. multiplicity of failure modes is the primary behavior of

a ship hull. In the primary behavior, one treats the ship as a single beam ~
subjected to weight, buoyancy and wave loads which induce sagging and hogging
bending moments. The hull may fail (or reach a limit state) in one of several
possible modes, e.g., buckling of deck or bottom panels or grillages, yielding of
deck or bottom plating, etc. Here again, system reliability methods must be used
to combine these different modes of failure and to obtain a total probability of
failure.

Multiple, modes of failure of a member are usually modelled in system

reliability analysis as a series system. A seties svste m is one that is composed of
links connected in series such that the failure of any one or more of these links
J constitute a failure of the system, i.e., weakest li~ system. In the case of the
primary behavior of a ship hull, for example, ~nv ow of the failure modes
discussed earlier will constitute failure of the hull (or a limit state to be prevented)
and therefore can be considered as a series system. Series systems will be
discussed in more detail in a later section of this chapter.

The second source of system effects is due to redundancy in multi-

component engineering structures. In such structures, the failure of one
member or component does not constitute failure of the entire system. Usually
several members must fail to form a failure path before the entire structure
ftils. The failure of each member is defied by at least one limit s-te equation
and a corresponding probability of failure. These individual member probabilities
of failure must be combined to get the probability of failure of the system for a
particular failure path. Thus, system reliability methods must be used to
determine the reliability of a redundant structure. An example of a multi-
component redundant stmcture in which system effects are important is a fixed
offshore platform. For such a platform to fail, several members must fail to form
a failure path. The probability of failure of the system in this case is usually
modelled as a pa rallel svsterq in which 911links along the failure path of the
system must fail for the entire structure to fail. More over, there will be several
possible paths of failure any of which will constitute failure of the entire platform.
Therefore each failure path and the associated probability of failure can be
considered as a link in a series system since failure of any link constitute a failure
of the system in the series model. The total offshore platfomn can be thus
modelled as several parallel subsystems each of which represents a failure path
connected together in series since any of them constitutes failure of the platform.
Parallel systems and general systems consisting of series and parallel
subsystems will be discussed in later sections of this chapter.

8.2. General Formulation

The exact system reliability problem taken into consideration possible time-
dependent random variables is an outcrossing problem. If the time-dependant
loads or response of the structure exceeds (outcrosses) one or more of several
possible failure modes (surfaces), failure of the structure occurs. The problem
formulated in terms of stochastic processes however is difficult to solve. Only a
few cases of very simply structures with certain load history models can be

196

Zog
evaluated in this manner and the reliability of the stn.mture at any time during its
life can be calculated. For a single time varying load it is possible to treat the
peaks as a random vaxiable and its extreme-value distribution may be formulated
to perform the reliability calculation.

At the present, the general problem is formulated as a time-independent

problem which is suffwient only for the evaluation of an instantaneous reliability.
As such, the form of the equation to evaluate the system reliability is the same as
that cifcomponent reliabili~ (equations 4.1 and 4.2) except that, now, the multiple
integration is earned out over all possible limit state functions corresponding to
the potential modes of failure. For k modes of failure, and n random variables,
the system probability of failure can be written as:

Pf= ~ ... J fx(xl, xz, . ..xn) ~l...dXn (8.1)

gi@<O
i=l,2, . ..k

where fX (xl, X2 . . . xn) is the joint probability density function of the n random
variables and gi(~ are the k limit state fuctions. The domain of integration in
, equation (8.1) is over the entire space where each of the k limit state function is
negative or
. zero,

The same difficulties encountered in the level 3 computation of component

reliability will be encountered in determining system reliability from equation
(8. 1), namely, the determination of the joint densi@ function and the evaluation of
the multiple integration. In addition, the domain of integration over all possible
modes of failure in equation (8.1) will present additional numerical difficulties.
For these reasons this general exact formulation is not used, and instead of
determining the combined total probability of failure of the system as given by (8.1)
only an upper and lower bounds on that system probability are determined. These
upper and lower bounds are usually determined by considering the structure to be
a series system or a parallel system or a combination of both (general system).

197
 It should be noted that, in principle, simulation methods and the Monte
Carlo technique can be used to solve equation (8.1) in basically the same manner
discussed in Chapter 7. In this case numerical simulation of the random
variables is performed according to their prescribed joint distribution and all limit
state equations are checked to see if failure occurs. The ratio of failure
realizations to tital number of simulations gives an estimate of the probability of
system failure. Reduced variate techniques and other methods for improving
convergence may be used here. UsUally, for realistic stmctures the number of
simulations required for a reliable estimate of the system probability of failure is
still high, but these methods have potential for application in system reliability.

.
8,3. Bounds on the Pro babilitv of Failure of a Series SVS~

A series system is one which fails if any one or more of its components
fails. Such a system has no redundancy and is also known as weakest link
system. Schematically a series system is represented as in figure 8.1

----m--m+ 3t---
Figure 8.1. Schematic Representation of a Series System

A typical example of a series system is a statically determinate structure

where a failure of any member constitutes failure of the structure. Another
example of a series system is a beam or an element which may fail in any of
several possible modes of failure each of which may depend on the loading
condition of the beam. A ship hull girder in its primary behavior is such a
system with the additional complication that failure may occur in hogging or
sagging condition. Each condition includes several modes of failure. A third
example of a series system arise when combining the probabilities of failure of
several possible failure paths in an offshore platform, any of which constitutes
failure of the platform.

If Fi denotes the i~ event of failure, i.e., the event that [gi W q 01, ~d Si
represents the corresponding safe event, i.e., [gi ~) a 01, then the combined

19Q
system failure event FS is determined as the union U of dl individual failure
events Fi as

Fs =UFi i=l,2, . ..k

1

The corresponding probability of system failure is

+(y Fi) =
P(Fs) 1- P
()
nsi
i
(8.2)

where n represents the intersection or mutual occurrence of events.

The calculation of the probability of systems failure for a series system

using equation (8.2) is generally difficult and requires information on correlation
of all failure events. Approximations are therefore necessary and upper and
lower bounds on the system probability of failure are constmcted instead of
evaluating the exact value. Two types of bounds can be constructed; first and
second order bounds.

First Order Bou~

These are bounds on the probability of system failure which require no

information on the correlation between the events of failure. In other words, the
user of such bounds does not need any information on the correlation between the
events of failure which, in many cases, are not available. They are constructed as
follows (see reference 8.1).

If the events of failure of a series system are assumed to be perfectly

correlated, the probability y of system failure is simply the maximum of the
individual probabilities of failure. For positively correlated failure events, this
assumption leads to the lower non-conservative bound on the actual system
probability, i.e.,

max. P (Fi) s P (Fs) (8.3)

1

199

211
 On the other hand, if the events of failure are assumed to be statistically
independent, an upper bound (consemative) can be detemnined. In this case, for
independent, failure events of a series system, the right h~d side of equation (8.2)
reduces to

k k
1-P nsi =1- ~ P(Si)=l - II [1- P(Fi)l (8.4)
() i i=l i=l
.
k
where 11P( Si ) represents the product of the probabilities of suwival. The result
i= 1
given by equation (8.4) represents an upper bound on the true probabili~ of system
failure, i.e.,

k
P(Fs)~ 1- II [1- P(Fi)] (8.5)
i=l

Combining equations (8.3) and (8.5), one obtains an upper and lower
bounds, i.e.,

k
max. P(Fi)~P(Fs)~ 1- II [1- P(l?i)l (8.6)
i i=l

Although the upper bound in equation (8.6) is not dMcult to evaluate, it can
be further simplified by noticing that

k k
1- 11 [1- P(Fi)l s X P(Fi) (8.7)
i=l i=l

therefore, equation (8.6) can be written as

k
max. P(Fi) ~ P(Fs) ~ z P(Fi) (8.8)
i i=l

200
 Equation (8.8) gives the final result for the bounds of a series system and
states the obvious conclusion that the actual probability of system ftiure lies
between the maximum of the individual probabilities and the sum of all
individual probabilities. These bounds are narrow if one mode of failure is
dominant, i.e., if one of the individual probabilities of failure is much larger than
the others. If not, these bounds may be tio wide to be useful. In such cases a
more narrow set of bounds should be considered (second order bounds).

s~co nd Orde r Bounds:

These bounds were developed in references [8.2, 8.3, 8.4 and 8.5] and are
given in terms of pair-wise dependence between failure events, therefore, are
called second order bounds. The original bounds for k potential modes of ftilure
are given as [8.2, 8.31:

k i-1
P(F1) + z max. { [ P(Fi) - E P(Fi Fj) 1 ; O} ~ P(Fs) s
i=2 j=l
k, k-
Z P(Fi) - Z max. P(Fi Fj) (8.9)
i=l i=2 j<i

where P(F1) is the maximum of the individual probabilities of failure ~d P(Fi Fj)
is the probability of inte~section (mutual occurrence) of two events of failure, Fi
~d Fj. P

. .

The bounds given by equation (8.9) depend on the ordering of the failure
modes and different ordering may correspond to wider or narrower bounds.
Therefore, bounds corresponding to different ordering may have to be evaluated to
detmn.ine the narrowest bounds.

The evaluation of the joint probability P(Fi Fj) required in equation (8.9)
remains difficult. A weakened version of the of these bounds (more relaxed
bounds) was proposed by Ditlevsen in [8.4] as follows.

In the lower bound of equation (8.9), P(Fi Fj) is replaced by [8.5]

201
 P(Fi Fj) s P(A) + P(B) (8.10)
.
whereas, in the upper bound, the same term is replaced by

P(Fi Fj) = max [ P(A), P(B) 1 (8.li)

where

(8.12)

(8.13)

and W) is the standard normal cumulative distribution fmction and pi are the
individual safety indices (Hasofar-Lind) as discussed in Chapter 5. p is the
correlation coefficient between two failure events (or modes). Such a correlation
coefficient between the failure events (Fi) = (gi ~s O) ~d (Fj) = (gj N ~ O) - be

evaluated from [8.4]:

(8.14)

where

Sg.

[1[1
agj
COV (gi, gj) = mgl#m. 8xrn* (8.15)

and
gi
= [511
[21 (8.16)

(8.17)

202
 In equations .(8.14) to (8.17), ~, ~ ., .% are the reduced random variables
and the derivatives are evaluated at the most likely failure points as discussed in
Chapter 5. The proposed bound by Ditlevsen [8.4] apply only for normaliY
distributed random variables.

Narrower bounds than the second order bounds can be constructed, but,
they involve intersection of more than two failure events and are much more
complicated.

8.4 Bounds on the Proba bilitv of Failure of a Parallel Svstem~.

A parallel system is one which fails only if all its components fail, i.e.,
failure of one component only will not necessarily constitute ftilure of the system.
Schematically, such a system can be represented as shown in Figure 8.2

IZIEIEI

+
Figure 8.2. Schematic Representation of a Parallel System.

A typical example of a parallel system is a statically indeterminate

structure where, because of redundancy, failure of several members along a
failure path must take place for the entire structure to fail. The behavior of
such a structure depends also on whether the members are brittle or ductile.

203
 Generally, brittle failure implies that the member looses completely its load-
carrying capacity while in ductile failure, the member maintains a certain level
of load-carrying capacity after failure.

If Fi denotes again the i~ event of failure and & the corresponding safe
event, then the system failure event of a parallel system Fp of k components (i.e.,
failure events) is the intersection or mutual occurrence of all failure events, i.e.,

Fp = flFi i=l,2, . ..k (8.18)

i

The corresponding probability of system failure is

P(Fp) =PnFi (8.19)

() i = 1-p(Ysi)

Equation (8.19) for failure of a parallel system should be compared with

equation (8.2) for failure of a series system. It is clear that the. failure of a series
system is the union (any) of the component failures, whereas, the failure of a
parallel system is the intersection (all) of the component failures.

Just as in a series system, the evaluation of equation (8.19) for determining

the exact system failure of a parallel system is generally difficult, and,
approximation by constructing bounds is usualIy necessary.

Simple first order lower and upper bounds can be constructed using
similar arguments as for the series system. Now however, perfect correlation
between all failure events (p = 1.0) corresponds to the upper bound and no
correlation between any pair corresponds to the lower bound. Thus, for positively
correlated failure events, these bounds are:

k
II P(Fi) s P(FP) & @n. P(FJ (8.20)
i=l 1

Unfortunately, the bounds given by equation (8.20) on the probability of

failure of a parallel system are wide and no second order bounds are available. In

204
some special cases, however, the exact system failure can be evaluated. For
example, Theft-Christensen and Baker (see reference [8.7]) evaluated the
probability of parallel system failure under deterministic loading and other
restrictive conditions.

8.5 General Svst ems,

A general system is one that consists of a combination of series and parallel

subsystems. A useful general system from an application point of view, is one
that consists of parallel subsystems connected together in a series. An example of
application for such a general system is an offshore platform (or, in general a
statically indeterminate stmcture) where each failure path can be modelled as a
parallel subsystem and all possible failure paths (parallel subsystems) are
connected together in a series since any of them constitute failure of the platform.
This representation is called minimal cut set since no component failure event
in the parallel subsystem (a failure path) can be excluded without changing the
stati of the structure from failure to safe. A schematic representation of parallel
subsystems connected together in a series is shown in figure (8.3).

Fm- ~F31

%
--m-
Figure 8.3. Schematic Representation of Parallel Subsystems
Connected in a Series (Minimal Cut Set)

A general system may also consist of a series of subsystems connected

together in parallel (minimal link set). Such systems, however, have less
potential for application to structural reliability and therefore will not be
discussed further.

205
 The failure event Fg of a genenl system consisting of parallel subsystems
connected together in a series (minimal cut set) is given by the union (seties) of
intersection (parallel) of individual failure events, i.e.,

Fg=j un
i (Fij) (8.21)

where (Fij) is the iti component failure in the ~ failure path. The probability of
failure of such a system is thus determined from

P(Fg) = P[y n
~ (Fij)
1 (8.22)

Exact evaluation of (8.22) is difmilt and requires information of the joint

dependence of failure events. Similarly bounds on the probability of failure given
by (8.22) are not available in general. If however, one is able to determine the
probability of failure of each parallel subsystem (for example, under restrictive
conditions), then first or second order bounds can be determined using equations
(8.8) or (8.9) for the remaining series system.

8.6 The Probabilistic Netwo rk Evaluat ion Techniaue (PNET);

The PNET is an approximate method for estimating a single value of the

system probabilityy of failure rather than bounding it [8.5]. The motivation behind
the method is that the bounds given by equations (8,8) or (8.9) for a series system
can be wide if none of the events or modes of failure is dominant, i.e., if none of the
probabilities of failure is much larger than the others. The same problem is
encountered if the series system consists of parallel subsystems (failure paths)
none of which is a dominant failure mode.

The PNET method is based on the fact that perfectly correlated (or, as
approximation, highly correlated) events of failure in a series system have a
system probability given by the lower bound of equation (8.6) or (8.8), whereas,
independent failure events have a system probability given by the upper bound.
Therefore, one may select a threshold value for the correlation coefficient and

206
assumes that failure events with correlation coefficient ~lj above or equal ta the
threshold value to be perfectly correlated, thus can be represented by a
representative event which is the event among them that gives the maximum
probability of failure, say P(Fr) (see lower bound of equation 8.8). M a set of n
failure events (modes) Fi where i = L . . . n is amanged in a decreasing order then
the failure probability of the representative event is P(F1). The remaining, events
with p Ij c p ~ are again rearranged in a decreasing order of their failure
probabilities. Let these be F2, F3, . . . Fk and the pair-wise correlation are p&, p&,
. . . P&- Those ev@nts with p~j >0 are represented by F;. The rem~~ng ones are
rearranged in a decreasing order and the procedure is repeated to search for
other representative events (modes) of failure. The mutual correlation between
the representative events will be low therefore, they may be assumed independent.
Thus, the probability of system failure maybe approximated by (see upper bound
of equation 8.6)

P(F) =l-r n [l- P(F,)] (8.23)

8.7 Fault Tree and l%e nt Tree

Fault tree is a systarnatic and effective method of identifying various

possible failure events and their interaction that lead to a main failure event
called top event. It is usually represented by a fault tree diagram which
starts with basic events whose probabilities of occurrence may be readily
estimated and describes the various possible combinations (unions and
intersections) of such events that lead to the top event or failure of the system. Its
value becomes more important in complex systems where some possible modes of
failure may be overlooked.

Fault tree analysis finds many applications in the design and operation of
nuclear power plants. It can also be applied to complex stmctural system such as
offshore platforms. In such analysis the fault tree diagram will help in
identifying in a systematic manner the various component failures that form a
failure path and the different possible failure paths that will lead to the top
event, the failure of the entire structure. In addition to identifying all potential

207
failure paths, the fault tree analysis may single out the critical events that
contribute significantly to the likelihood of failure of the structure.

The probability of top event (main failure) in a fault tree analysis is

calculated through unions and intersections of subevents which are expressed in
terms of basic events (component failure) for which the failure probabilities can be
estimated. Although fault tree analysis provides the logic leading to the top event,
it does not eliminate the dif3kuIties in computing the probabilities of unions and
intersections of comelated events. Approximations and bounds may still have to
be used if a quantitative assessment is to be made of the probability of occurrence
of the tap event.

Qualitative evaluation of a fault tree can provide also valuable information

to desigpers. Without knowing accurately the probabilities of failure of events and
subevents, the fault tree analysis may point out the critical basic events and the
critical paths that contribute significantly to the occurrence of the top event. With
this knowledge a designer may then take the appropriate steps to reduce the
probability of occurrence of such critical basic events.

Event tree analysis on the other hand starts with the top event (main
failure) and examines in a logical manner all possible consequences resulting
from the occurrence of such an event (for example loss of life, pollution, explosion,
fire, etc). The consequences of the top event (now called the initiating event) may
or may not be series depending on the possible occurrence of other adverse events
following the initiating event. The identdlcation of all possible subsequences and
scenarios is best accomplished through an event tree diagram. Each path in
the event tree represents a sequence of subsequent events leading ta a particular
consequence. The probability of occurrence of a spetic consequence depends on
the probabilities of the subsequent events and is simply the product of conditional
probabilities of all events along that path.

8.8 Reliability Bounds for shin Primarv ~tren~~

Reliability bounds for ship primary strength were developed in 1972 in

reference [8.6]. In the primary behavior, the ship hull is considered as free-free
non-uniform beam supported by water pressure. Wave loads (bending moment)

208

220
are calculated using the equations of motion of the ship if dynamic effects are to be
included otherwise by balancing the vessel on a wave configuration. The loads on
the vessel alternate from hogging which produce compression in the bottom
plating to sagging which induce compression in the deck. This hog/sag variation
must be considered in the hull reliability analysis.

In each hog/sag condition there will be several possible modes of failure,

eg., plate and panel buckling, tensile yield, etc. If Fh and FS represent hogging
and sagging events of failure, respectively, then the combined event of failure FCis
given by the union of the two events, i.e.,

(Fe) = @h) U(FS) (8.24)

Since hogging and sagging are mutually exclusive events, i.e., the vessel
can be either in hogging or in sagging condition (but not both at the same time),
then the union of the two events given in (8.24) is simply their sum. The
probability of combined event of failure is thus

P (Fc) = P (Fh) + P (Fs) (8.25)

As mentioned earlier, each of the hogging and sagging conditions will have
several possible modes of failure (or limit states). In each case these modes can be
modelled as a series system since any of them constitute a failure of the hull (or a
limit state to be prevented). Thus bounds on the probability of failure in hogging
condition P(Fh) and in sagging condition P(Fs) can be constructed using equation
(8.8) or (8.9) for first or second order bounds, respectively. The bounds on the
combined probability P(Fc) are simply the sum of the bounds on each condition as
implied by equation (8.25).

Experience indicated that in many cases either hogging or sagging

condition is governing in the. reliability analysis depending on whether the
stillwater bending moment is hogging or sagging (in much the same manner as
in the usual deterministic analysis). In some cases, however, both conditions
must be included othemvise the estimated reliability will be unconsemative.

209

221
8.8.1 Numen cal Exa mde~

In reference [8.6], the probabilities of fkilure (or reaching a limit stite) were
calculated for a Mariner class vessel in hogging and sagging conditions using
level 3 reliability method. The stillwater moment for the Mariner is a hogging
moment. Several modes were considered in each condition and the results are as
follows:

i) Tensile yield of deck plating pf = 6.16x 10-7

ii) Compressive post-buckling yield of inner bottom plating pf = 4.o3 x
105
iii) Compressive post-buckling yield of bottom plating under lateral and
inplane loads: pf = 62.99x 10-5
iv) Compressive grillage failure of bottom shell under combined loads:
Pf = 1,47x 105
The first order bound on the probability of fdure in hogging condition p?
were thus obtained (see equation 8.8) as

ma.x (6.16x 10-7,4.03x 10-5, 62.99x 10-5, 1.47x 10-s)< p?

< 6.16X l&T+ 4.03X 10-5+ 62.99X 10-s+ 1,47X 10-5
or
6.3 x 104 s p? s 6.8 X 10~ (8.26)

Notice that these bounds are tight since one mode of failure is dominant
(post-buckling yield of bottom plating). There is no need to consider second order
bounds.

b. Saztinr Condition

i) Tensile yield of bottom plating pf = 1.55x 10-14

ii) Inelastic buckling of deck plates between stiffeners: pf = 2.29x 10-11
iii) Grillage instability of deck: pf = 2.18x 10-7

210
 :

Bounds on the probability of failure in sagging condition p~ can be

,
constructed in a similar manner as those for the hogging condition. The sum of
the two sets of bounds would then give the bounds on the combined probability of
failure. It is clear, however, that in this case the bounds on the failure modes in
sagging are of the order - 10-7, much smaller than those given by equation (8.26)
for the hogging condition. The latter bounds, therefore, can be considered as
bounds on the combined probability of failure.

.
.-

211
 REFERENCES

8.1 Cornell, C. A., Bounds on the Reliability of Stmctural Systems, Journal

of Stmctural Division, ASCE, Vol. 93, No. STl, February 1967.

8.2 Kounias, E. G., Bounds for the Probability of a Union with Applications,
Annals of Math. Stat., Vol. 39, No. 6,1968.

8.3 Hunter, D., An Upper Bound for the Probability of a Union, Journal of
Applied Probabili@, Vol. 3, No. 3,1976, pp. 597-603.

8.4 Ditlevsen, 0., Narrow Reliability Bounds for Structural Systems, Journal
of Structural Mechanics, Vol. 7, No. 4, 1979, pp. 453-472.

8.5 Ang, A. H.-S. and Tang, W. H., Probability Concepts in Engineering

Planning and Design, Vol. 11, John Wiley and Sons, Inc., New York, 1984.

8.6 Mansour, A., Probabilistic Design Concepts in Ship Stmctural Safety and
Reliability, Trans. SNAME, Vol. 80, 1972, pp. 64-97.

~8.7 Theft-Christensen, P. and Baker, M., Structural Reliability Theory and Its
Application, Springer-Verlag, Berlin, 1982.

212
 9. FATIGUE RELIABILITY

9.1 INTRODUCTION

Fatigue is the degradation in material, element, and system strength and

stiffness as a result of cyclic straining-stressing.

Materials in marine structures can consist of steels, concretes,

synthetic fibers, and soils (foundation). Elements can range from

bulkheads to hatch cover openings, from cylindrical braces to mooring

lines, and from deep (piles) to shallow (anchors, mats) foundations.

Systems represent assemblages of elements, and can range from cargo ships

to fixed and floating platforms.

Cyclic straining can develop from a widevariety of environmental

(thermal, wind, wave, current, ice, earthquake), construction

(installation transport, launch), and operational (slamming, equipment,

cargo) causes. The relentless cyclic forces are perhaps one of the most

distinguishing characteristics of the marine structures environment.

Design for fatigue reliability has four p~incipal.lines.of defense:

1. Minimize stress-strain risers (stress concentrations) and cyclic

straining-stressing through good engineering of the structural

system and its details. This requires a high level of

engineering quality assurance (QA) at the concept-development-

design stage.

2. Minimize flaws (misalignments, poor materials, porosity-voids,

etc.) through good, practical material and fabrication

213

X2 !7
 specifications and practices. This requires a high level ofQA

during the development of plans and specifications and during

construction (involving materials selection, fabrication,

transportation and installation). Further, there is a similar

QA program required during operations to properly maintain the

system.

3. !!LW!l@ degradation at the local elemen~ through selection of

good materials and fabrication practices, and good engineering

designs (e.g. crack stoppers, damage, localizer, and repairable

elements). This requires a recognition that when (not if)

fatigue degradation occurs, all reasonable precautions are taken

to restrict its development and effects. Note, again QA plays a

key role, particularly during operations to disclose the

presence of fatigue degradation (early warning). -

4. Minimize ~a~~he svstem~ so that when (not if)

local fatigue degradation occurs, there are no significant

effects on the systems ability to perform satisfactorily. Here

good fatigue design requires system robustness (redundancy,
ductility, capacity) and system QA. Inspections and monitoring

to disclose global system degradation are another strategy to

minimize potential fatigue effects.

The purpose of this discussion has been to outline the major factors and

the complex interplay of these factors in determining fatigue

reliability. Cyclic strains, material characteristics, engineering

design, specifications, and life-cycle QA (inspections, monitoring) are

all parts of the fatigue equation. This is the engineering equation of

fail safe design -- fatigue may occur, but the structure can continue

214

22 b
to function until the fatigue symptoms are detected and repairs are made.

The alternative is safe life design - no significant degradation will

occur and no repairs will be necessary. Safe life designs are difficult

to realize in many long-life marine structures or elements of these

structures.

Uncertainties and variabilities are present in each of the parts, and

thus, reliability methods can play an important role in assisting the

engineer to achieve fatigue reliable-durable structural systems.

215
9.2 FATIGUE
ANALYSIS

A fatigue analysis can be organized into five basic components:

1. Characterize the life-cycle (short term and long term) CYC1ic

conditions.

2. Determine the cyclic forces imposed on or induced in the

structure (system).

3. Evaluate the cyclic strains-stresses developed in the element

(detail) of concern.

4. Determine the degradation in strength and stiffness (damage of

the element (detail) caused by the cyclic strains-stresses).

5. Given the fatigue damage, evaluate the acceptability of the

element (detail) performance.

-
In development of the following simplified fatigue design procedures

[?.1], itwill be assumed that waves are the source of cyclic forces. It
will be assumed that the long-terni(e.g. T = 100 years) wave height

distribution can be represented by Weibull distribution (Figure 9.1).

For storms, the cumulative distribution function [CDF = FX(X)] of wave

heights (h) is:

(9.1)
FH(h) bexp[.(;)k]

216

22t
 Hunkane Cornponem
60 -
(g, )=l)

50
.
40 <0
\ \
\ \
30- b,
I \ \
20
O~rational
\
\
lo-
\
\

0 t t 1 i i
1 to 100 103 104 105 106 j(p @
w

-z N - Number of Waves Exceeding h (cycles)

FIG. ~.1 TWO-COMPONENT WAVE HEIGHT DtSTRIBUTION

For non-storm conditions, the CDF is:

FH,h, = l-exp[-($]mo] (9.2)

The structural detail fatigue stress range (peak to peak) (Sf) will be

taken to be a function of the wave height:

s, = CHa (9.3)

Next, the number of cycles to failure (N) of the detail subjected to a

cyclic stress range (Sf) will be taken as (Figure 9.2)* [9.2]:

N = KS;m (9.4)

Accumulation
... of fatigue damage (D) will beassumed to be described by a

linear damage accumulation rule (Palmgren-Miner):

(9.5)
R(Sfi)
D= I1 N(Sfi)

*Footnote:

Log N = Log K - m Leg Sf

and, Sf= (K/N)l/m

and, N = K Sf-m

m = negative slope of S-N curve

Log K = life intercept of S-N curve

218
 TOTAL CYCLIC STRESS RANGE (ksi)
- 0.s 1 2 5 10 20 so Im
11 1[ I I
I I I I
I 111 I 111

o* 1,,,,1 1 1 I 1,,,,1 1 I 1
50 100 200 500 1000 2003 5000
HOT SPOT STRAIN RANGE X II)

CURVES X AND X

219
Where

n(Sfi) = number of stress cycles at stress fi

N(Sfj) = number of cycles to failure at stress fi.

The summation is overall stresses, Sfi, experienced by the structural

detail. When D = 1, failure is presumed to occur.

Fatigue damage (DL) accumulated over the life (T) of the detail can be

computed from the following equation [9.1]:

( 9.6)
DL = *( YO+YJ

Where

-s (9.7)
. m
No am
Y. = ~~~m(lnNO)fO~
()
1+Eo

- am ( 9.8)
am
N]
Y, = ~H~(lnlJl)~
( )
l+
El

r(a) = Gamma function

Now, let the design accumulated damage be limited to a fraction of the

life damage:

220

232
 DL (9.9)
D~=
F.,

or the design service life (T~) be:

T. = (F,,)OT (9.10)

where

F~f = fatigue life or damage Factor of Safety (commonly in

the range of2 to 3).

The fatigue design stress (SfD) will be related to the fatigue design

wave height (HfD) as before:

(9.11)
s ID = CH;D

Thus,

(9.12)

s ID =
T,(YO+YJ

Based on an HfD = 70 feet,* the long-term wave height distribution in

Figure 9.1, and the API-X S-N curve (weld without profile control, Figure

9.2), SfD can be computed as functions ofT~ andm (Figure9.3) [9.1].

Note:

Use an HfD close to extreme condition design wave height, e.g. 100-year
height.

221
222
For example, for a detail life of 25years and a factor-of-safety of 2

(TS = 50 years), and a wave height exponent (a) of 1.0, the design stress

range is indicated to be 65 ksi. For a detail stress concentration

factor (SCF) of 3, this would equate to a nominal allowable design stress

of 22 ksi (based on API X S-BN curves, Figure 9.2). For an a= 1.5, the

design stress range would increase to 150 ksi. For an SCF = 3, the
nominal stress range would be 50 ksi.

In design practice, it is often useful to state the design stress range,

SfD, as a peak stress value, SpD. This can be accomplished by defining a

stress ratio, R, that is the ratio of the minimum stress, Smin, to the

maximum stress, Smax for the design wave height, HfD. Thus:

SjD (9 .13)
s PD =
(l-R)

The values of Rwill be structure dependent. For conventional

template-type, shallow water platforms, R is typ~cally close to -0.33.
,-

? .

223
9.3 RELIABILITYllODELS

Based on a fatigue analysis as outlined in the previous section, the

principal sources of uncertainty can be organized as follows:

1. S-N relationship (Eq. 9.4, Figure 9.2)

2. D-N relationship (Eq. 9.5)

3. S-H relationship (Eq. 9.3)

4. H-N relationship (Eqs. 9.1, 9.2)

5* S-SCF relationship

6. DD-DL (or LD-L) relationship (Eqs. 9.9, 9.10)

7. Other factors such as corrosion and cathodic protection

Additional details on these sources of uncertainty are given in Table 9.1

[9.3]. There are many sources of complexly interrelated uncertainties

and variabilities. It is the primary purpose of a fatigue reliability

analysis to logically organize these sources, and then to quantitatively

evaluate them to determine what factors-of-safety (e.g. Eqs. 9.9 and

9.10) (alternatively, levels of reliability) should be employed in a

given design-analysis framework.

Wirsching [9.3, 9,4] has made extensive fatigue reliability studies for

fixed offshore structures. Munse [9.5] has made similar studies for ship

structures. Recently, Wirsching and Chen [9.6] have summarized and

contrasted results of these two studies. The following discussion will

be based on these developments.

Alternative methods for computing fatigue damage are summarized in Figure

9*4 [9.3]. Table 9.2 [9.3] summarizes the alternative analytical

expression for computing fatigue damage at a joint (detail).

224
 TABLE q.1

A SUMMARY OF FACTORS AND CONSIDERATIONS

RELATED TO FATIGUE IN WELDED JOINTS OF OFFSHORE PLATFORMS

ATIGUEBEHAVIOR
OF WELDED
JOINTS
Oeffnftfon of f~tlgue faflunfnS-Ndata
Size effect in S-N datS
Effect of ~ld pruffle
. Effect ofcormslon ard cathodic pmtectl~
Assumtlon of llnearmdel md lognomsl distribution for N
Classiffcatlon of jofnt on the bas~s of gemmtm rlther than load Ptttem
. Relationship ktwen stress at jointandstress used to obtain S-N curve
Ignoring possiblestnss endurance in 5-NCUIWC
Cmqatibility of determination of hot SpOt stress with S-N CU~C

WNUFACTURIN$
CONSIDERATIONS
9 F~bri cation uncsrtaintf es
Requi mmnts on wld contmm notmt
DEFINITION OF THE ENVIROWENT
s Use of full scitter diagram of ti~ - To
Variations in TB
Z occummceestimates
Wavedirectionality
Interaction of Waves and Currents
Theo~eti cal mdel used for ocean waves
HYDRODYNAMIC
LOAbSON STRUCTURE
Inertia and Drag Coefficient
Oimctional wave spectra uhich accounts for wave Spreadf ng
Marine growth
Sheltering effects
STRUCTURAL
RESPONSE
TO tiYMODYWIIC LOADS
. Ass@tions mde In spectml analysis
a. 1i near mponse dud ng transfer functf m devehmmnt.
b. 1t nearf zatlon of drag tem
c. at jofnts, 1. No flexibility
f1.ffectofan
iif.centertocenter
cmrdf nates
d. Sof 1 stf ffness in Dynamic Mo41
. D@ ng *f fects f n stntctiral response
f. Dynamic res~se mt accounted for fnanalysis
FATIGuE
STRESSES
ATJOINT
Hsthod of slmis to vtlwte stress concmtrstion
factors
{SCF)
. Paranmtrfc quatfm used for SCF
. Pofnt at lnte~eei~whitefaf
llt~0CCW5
FATIGuEDAMAGE
EWTIONS
. Assmption of Nfners Rule
Am,smtlon of nerrw bsnd @sage equation
fn spectral approach
Ammqtian of Uefbull di stributf on for stress ranges in stress distribution
pproach
WHERCONSIDERATIONS

. Bad judgenent during Wfng and fns~l latfon

In semica lads
 FIG. %4 cLAssfF~CATION OF FOUR BASIC METHODS OF COMPUTING FATIGUE DAMAGE

Definition ~

w
M
M Have Melght Histo;y Spectral Analysis
u-l
1
I Stress Range History
I

* v T

Have Exceedance Wave distribution 1 Stress Range Distri- The Equivalent

Spectral Method
Histogram (some- method; stress assuned bution Method; stress Weibull Method:
(sometimes called
tfmes called the proportional to wave range model led theore- the probabilistic This method used
determlnlstic hetght, and wave tically, usually by OnV, An
method). Damage
method). This height distribution Weibull. Stress range equivant Weibull
computed from each
Lechnique used by described by lognonnal, distribution can be seastate fn scatter distribution of
Lloyds. Damage lle~bull,etc. Nolte- derived from spectral long term stress
d~agram, in which
computed from Hansford methdo method, wave distri- ranges is derived
stress is consider-
wave exccedance assunes We!bull distri- bution method, ZLC. using spectral
ed a stationary
diagram bution. analysis
random l}rocess

N
Q-d

%7
 TABLE %2 ~

A SUMMARY OF EXPRESSIONS FOR FATIGUE DAMAGE JOINT

FATIGUEDAMAGEAT TIME T,

D = TBmQ/K

m,K = paramters fran S-N curve

B = factor to account for uncertainties in estimating fatigue

stresses from oceanographic data

n = stress parameter ~

STRESS PARAMHER USING VARIOUS APPROACHESTO THE STRESS OISTRIWJTION

Wave Exceedance Diagram (Deterministic Method) .

n= fozcisim
i
f. = average frequency of stresses

si = stress range
Gi = fractionof total stress ranges that Si is acting

Spectral Method (Probabilistic Fiethod)

n= k(m)(247)mr(f+ l)~Yifiuim

X(m) = rainflow correction

r(-) = gama function

Yi = fraction of time in ith seastate

fi frequency of wave loading in the ith seastate

. th
ai = RMS of stress process in the I seas Q te

Heibull Model for StressRanges

Wirschings reliability analysis has been cast in a lognormal format in

which the random variables are assumed to have lognormal distributions.

The time for fatigue failure (T) is expressed as a function of the

accumulated damage (D = A), the S-N curve parameters (K, m, Eq. 9 .4), a

stress range model error parameter (B = actual/computed stress range),

and a stress range parameter (n, Table 9.2):

(9.14)
,.fi
Bf2

The probability of a fatigue failure (Pf) is taken as:

Pj = P[TST,] (9.15)

where P[] is the CDF of T, and T~ is the service life.

P, = @(-p) (9.16)

where *c] = standard Normal distribution function and P is the Safety

Index:

ln~/T. (9.17)
P ~,n7

where
p. = median value of T, and

Note: z will be taken as the median (50th percentile)

value of the parameter x.

228
 T==- ( 9.18)
Bmfl

U,nT = [ln(l+cj)(l+c~)
(l+c~)m] (9.19)

in which the Cs are the Coefficients of Variation (COV = C) of each

variable.

Uncertainty in the stress ranges (S) is expressed through the stress

range model error parameter (B). The errors are attributed to:

1. Fabrication and assembly (BM)

2. Seastate characterization (Bs)

3* Wave load predictions (BF)

4. Determin~tion of member loads (BN)

,---

5. Estimation of stress concentration factors (BH)

Thus,

B = BMB.B~BN.B~ (9.20)

and

c,= [(?(,+ci)-q (9.21)

for i =M, S, F, N, H.

229
Tables 9.3 and 9.4 summarize the statistical estimates on the B

components and K, B and d random variables, respectively. At the bottom

of Table 4, note the range of Safety Index implied by APIs RP 2A design

wave peak stress rule (limits nominal brace stress to 60 ksi).

Munses fatigue reliability analysis [9.5, 9.7] has been based on a two

parameter Weibull distribution of stresses (S) and cycles (N) (Figure

9*5).

1!
(9.22)
F~(s) = P(S<S) = l-exp
[()]
- ~

where

E = Weibull distribution shape parameter

6 = Weibull distribution scale parameter

Defining a design stress (Sm) such that this value is exceeded on an

average once every NT times:

(9.23)
~= Sm[llliVT]:

NTA is the total number of cycles in the service life T. Thus,

(9.24)
F,(s) = l.exp[-(:)flm,]

*Note: NT = Ts f. = Service Life x Average Stress Frequency

230
 TABLE %3

SUMMARY OF BIAS AND COEFFICIENT

OF VARIATION OF COMPONENTS OF 6

Random variables representing sources

of unwk.inty infatiguestress estimates Bias Cov
(1) (2) (3)
BM 0.90-1.W O.lo-o.w
Bs o.6&-l.20 0.4W0.W
BF O.w-l.lo 0.10-0.30
BN 0.80-1.10 0.20-0.0
l?~ 0. W-1.20 O.lo-o.w
Bias = actual load or stress/load or stms estimated by current analysis pro-
cedure; for each Bi, the bias =n be interpreted as the median value 5,.
COV = vexp (az) - 1, in which u = h (Xu/XJ/6; and XU and XL = upper
and lower limits of X.
CBS= FPBn = (0.9)(0.7) ~ 0.60 in which P = percent occumence of each seas-
tate; and D = directiondity.
This relatively large figure, which dominates C-, is due to the sensitivity of
the dynamic response to small variations in Tn. The figure of 0.40 is due to this
effect only.
1hh figure was obtained by ~.218 by assuming maximum COVS of 0.4
for dynamic response, and 0.3 each for directionalityand percent occurrence ef-
fects. Theresultingfigureof 0.60 is mnsidemdto be the largest reasomble vake.
This bias occurs when wave spreadingis not considered in the development
of theresponse spectra.
These extreme values should be used cmlywhen supportingevidence exists.

231
 TABLE %4

EVALUATION OF SAFEIV INDEX, $

IMPLIED BY API RP2A DESIGN WAVE PEAK STRESS RULE

USING DIFFERENT DATA SETS

I Data$et
Daslgnfactom AIB CID E F
[1) (6) (7)
-
S-N ame, inkips m 4*W 4.38 4.42 3*ocr 3.ti 3.00
per squm in~ K 4.6E12 4.6E12 L55F12 5.25E1O 1.29E11 1.46E1O
units . c~ 0.73 O*73 1.35 0.73 1.25 0.67
Rainflowcorrection k 0.79 0.79 0.79 0.86 0.85 0.86
Damage mtio A 1.(M) 1.00 1.00 1*(K) i.oo 1.00
c 0.30 0.30 ,0.30 0.30 0.30 0.30
Stress modeling BA O.w o.7d o*7cf 0.7(Y 0.7CY o.7cf
error c~ 0.17 0.50 0.50 0.50 0.50 0.50
Average frequen~, f., I
in hertz 0.25 0.25 0.25 0.25 0.25 0.25
Safety index implied bys
RP2A design wave peak
stress de (60 .ksi rl.lk)i B 5.34h 12.78 2.09
1
1262
2.57 1.83
Data from Conunentasy of API RF2A, ~. 81, Fig. (2.5.3-2
AWSX data, elastic range only.
T and K joint data providedbymemberof TechnicalAdvisoryCommittee
T and K joint data: an improved vemionof data set D.
VaIues provided by memk of Technical Advisory Committee. Value of C~
now thought to be 10W;
Values provided by memberof TechnicalAclvisory Committee. Numbers are
now considered reasonable for %orst ~ analysis in which wave spreading
and wave directionalityare not mnsidered.
As computed by solving for ~ in Eq. 7.16
Relatively high value due to srmdlvakeof C.. See supersaipt e above.
~~e T~e from UK K)C)ERULES
Note: For a 2@year life, S~ = 53.2 ks~ ~ = 0.69 for 20 year-wave. (Same stmc-
tum would have S~ = lUi ksi and ~ = 0.57 for I(Byear wave climate.)

232
Figure 9.5 [9.3] shows this distributionfor NT = 108 cycles. Figure 9.6

[9.7] shows measured 1ong term, low frequency, wave-induced ship hul1

girder stresses. Shape parameters (~) in the range ofO.7 to 1.3 model
these data well.

The Weibull shape parameter (E) will depend on a large number of factors

such as wave conditions, type of structure, dynamic response of

structure, position of fatigue detail in the structure. E characterizes

the severity of the fatigue stresses relative to the extreme design

stress. ~ = 1 yields a straight line on a semi-log plot (Figure 9.1),

and F = 2 resultsin the Rayleigh distribution.

Guidelines for~ for platforms in the Gulf of Mexico are g ven in Figure

9.7 [9.8]. Waterline braces and floating marine structures may have ~ in

excess of 1.0. Munse reports [9.5] ES in the range ofO.7 to 1.3 for
hull girder stresses in tankers and cargo ships.

Munses fatigue reliability model addresses the uncertainty of fatigue

life (expressed as COV = $N) as a function of uncertainties in stress

evaluation (Cs), workmanship in fabrication of the details (Cc), and

fatigue assessment (cF): .

CN = [C;+my:+cyz (9.25)

where

CF = [S:N+C:R] (9.26)

233
 &
I I

r
I L I
1//
I
l/vfi

r I I I I I I I I I
Lq
e-
d
E

234
 .

4
k
g
a
I&l
> Kl=--ilii / fY .

. .

/4
.
.

//..

/
//.

/4
.

f-

m, o
o

235

.2y7
 DEPENDS ON
SHOAUNG DEPENDS ON
1.0 (SHELF PRoFILE ) NAT. PERIOD

VERY NARROW
/

O*5
BROAD

lkh GULF OF MEXICO

o! t I 1 I 1 * 1 I 1 t 1 [
20 50 100 200 400

a L
:
F f WATER 1
DEPTH ~ 100 H.
I I 1
w
o! I I 1 1 I I ! ! 1
& +~ MSL -20 -m -1oo -200
u
rm (B] POSITION IN STRUCTURE
FEET BELOW MSL

t DOMINATED
WATER DE~H ~ 100 FT.
o ! t
o I 2 3 4 5
(C) NATIJRAL PERIOO -
SECONDS.
AFTER REF. 9.8

FIG. 9.7 SUMMARY DESCRIP~N OF ~ FOR GULF OF MEXICO

FIXED PLATFORMS

236 Zjq
and

CSN = COV associated with S-Ndata

c~R = COV associated with Miners rule

m . slope of N-N curve

Based on available fatigue data [9.7], Munse estimates the following

coefficients of variation:

, CSN = 0.62

0 CMR = 0.15
0 Cs = 0.10
0 cc = 0.40
0 CN = 0.96

These estimates do not incluci,?

any effects due to corrosion. Munse [9.7]
recommends use of a total uncertainty (CN) of 0.80 until letter values

can be established. Mun~es CN can be directly compared with Wirschings

CK (0.73 to 1.35, Tables 9.1 and 9.4).

Again, assuming Miners rule and that D = 1,

(9.27)
E(sm) = ;

where R is the mean life, and E(Sm) is the mean or expected value of Sm,

gives the probability of failure (Pf) at the service life (N~) as:

237
 ~;l.ca
(9.28)
T, E(sm)r(l+c;lOs)
PI =
[
K
1
where

(9*29)
E(Sm) = A(mJSmF ~+ 1
()

Note that the Weibull shape parameter ~ has been approximated as:

E = c;108 (9.30)

and the scale parameter b determined as:

N (9.31)
b
= r(:+l)

where

N = mean life to failure obtained from a least squares analysis

of fatigue data.

As noted in Table 9.2, a correction factor, identified as the rainflow

correction factor Am [9.9] should be applied to the stress parameter

when the Weibull parameters (Et6) are based on an analysis which uses the
assumption of a narrow-band process (Rayleigh) used to describe short

term distributions of wave heights in each of the stationary seastates

composing the long-term distribution. ForS-N curve slopes ranging

between m = 3 and m = 4, and short-term seastate spectral widths greater

than O.5, ~m=0.86to~ =0.80 [9.3].

23s
9.4 DESIGNAPPLICA710NS
- DETAILS

Reliability evaluations of the fatigue design of a structural detail

(joint) can be developed from either the Wirsching lognormal based

distribution ofN (Eqs. 9.16 to 9.19) or the Munse Heibull based

distribution of N (Eqs. 9.25 to 9.29).

Using these two approaches, an allowable/design stress range, SfD, can be

defined based on a Weibull distribution of stresses as [9.5, 9.6]:

s fD = S,PR~ (9.32)

where

~f= Krn= mean stress range for failure at N cycfles ( 9.33)

()z

@ = +(:)1 ( 9.34)

= random load factor

(9.35)

= Munse Reliability Factor [9.7]

1
z1
(9.36)

1
z
R Fw =
E [ exp(~DulnT )

- Wirsching Reliability Factor [9.6]

239

2s--1
where Pfd and BD are the design probability of failure and safety index,

respectively.

Ideally, pfd andp, should be based on considerations of the design

details influence on capacity of the structural system, inspectability,

and repairability. Life cycle operations, reliability, and costs should

be optimized.

Munses fatigue design process proceeds through six steps:

Step 1 - Establish the expected loading history for the detail to be

designed. This is equivalent to choosing a Weibull distribution

shape parameter, F (Figure 9.5). ~ commonly ranges from 0.7 to 1.3

[9.7].

Step 2 - Identify the type of detail to be designed. An extensive

summary of typical ship details is given in reference [9.7].

Step 3 - Obtain the mean fatigue stress ranges and slope of the S-N

curve based on the type of details and the design number of cycles

(NT). Based on laboratory tests of typical ship details, reference

[ %7] summarizes Sf and m data.

Step 4 - Compute the random load factorv from Eq. 9.34 based on NT,
~ and m.

Step 5 - Compute an appropriate reliability factOr, RFM, from Eq.

9.35 based on an estimate of the COV of fatigue life (e.g. CN = 0.8)

and desired probability of failure (e.g., PfD = 10-2).

Step 6 - Determine the design fatigue stress range from Eq. 9.32.

24o

-&?
Wirschings [9.3, 9.4] procedure for a fatigue design check proceeds

through six steps:

Step 1 - Define an appropriate value of of the Weibull stress range

shapes parameter ~ (e.g. Figure 9.7) for the environment, type of

structure, and detail location.

Step 2 - Define an appropriate stress ratio (R = Smin/Smax, Eq. 9.13)

for the detail.

Step 3 - Define an appropriate S-N curve slope, m (e.g. Figure 9.2)

and then use Figure 9.8 to establish the design stress range,

SfD = SR, for the desired service life, Ts.

Step 4 - The peak stress value for the detail is computed as

SpD= SfD/(1-R) (Eq. 9.13).

Step 5 - The fatigue strength can be stated in terms of a nominal

stress by using an appropriate stress concentration factor (SCF) or

SpDN = SpD/SCF. :
, *

Step 6 - The detail (joint) is taken as satisfactory if SmS S~D,

where Sm is the hot spot stress corresponding to the design wave

(assumed to be the 100-year wave). Alternatively, SmN s SpDN, where

S~ is the nominal (brace) stress.

Wirschings procedure is based-on,a code calibration approach to define

the fatigue design safety index (P,) [9.3, 9.41. Table 9.5 summarizes

the data and method used to construct the design stress range curves of
Figure 9.8.

241

-2s2
 TABLE 9.5

SUHNARY OF THE DATA USEDTO ESTABLISH THE

DESIGN STRESS RANGECURVES OF FIGURE 9.8

Figure g.8(a) Figure 9.8(b)

m 4.38 3.00

k 4.6E12 5.25E1O

cK .73 0.73

h .79 .86

2 1.00 1.0

c, .3 .3

B 7 0.7

CB .5 0.5

FO (tiz) 0.25 0.25

P, 2.78 2.62

Method for constructing curves: Consider Data Set B for which the target
safety index is B. = 2.78. SR forTs = 100 years is computed directly using
Eq. (a) with~ = 2.78. The 20-year curve we know must pass through the
reference. From Eq. (a), B = 1.63 for the 100-year wave condition. Using
this 8, the 20-year curve is establish from Eq. (a). For#. = 2.78 for a
Ts = 40 years, Eq. (a) fixed SR=52.4 ksi. This value Is scaled to SR = 78.7
ksi for the 100-year wave. Then.corresponding to~ = 0.57, Eq. (a) gives @ =
2.11 for the 100-year wave. This value of~ in Eq. (a) is used to construct
the 40-year curve; a B = 2.11 for 100 years ensures B = 2.78 for 40 years.
Eq. (a) is:

SR(mS,D) = [ln(fo~J]: AK

[ A~OT~Bmexp(pOd,n,)~(~+ 1)
1

242
 0

U+a
-

DUO
00
z-
<
a

ma

I
Based on the fatigue analysis process outlined in Eqs. 9.1 through 9.13,
Geyer and Stahl [9.1] have developed a very useful simplified fatigue

design procedure. The key to this procedure is the use of a uniform

fatigue life criterion (Eqs. 9.9, 9.10).

Based on theAPI XS-N curve (Figure 9.2, m* 4.38), adeepwater Gulf of

Mexico wave height distribution (Figure 9.1), a design wave height of 70

feet, and a 20-year service life (T) with a factor of safety of2

(TS = 40-year design fatigue life), they developed the design fatigue
stress range curves shown in Figure g.9 as a function of the stress-wave

height exponent, Q (Eq. 9.11). For details and structures in which as

1.2, the low-cycle fatigue stresses developed by hurricanes has an

insignificant effect. However, forus 1.2, the low cycle hurricane

stresses can have a major influence on damage.

.

As an alternative to a design stress range, Wirsching and Chen [9.6] have

formulated the fatigue design as a design damage ratio (Eq. 9.5), d.:

yz (9.37)
AD =
13mexp(@Dalhr)

The safety check on the computed damage Do is:

(9.38)

where

T,f2 (9.39)
Do=
K

244
 320

280

240 lqN log N

200

and Hurricant

160

120

80

40

00 1 1 I I 1
2.4
.4 .8 1.2 1.6 2
Stress-Wave Ht Ex~ent (g)

FtG.%9 SENSITIVITY TO WAVE HEIGHT DISTRIBUTION

245
Using reasonable fatigue reference data for calculation ofdo as a
function ofp,, Wirsching and Chen developed Figure 9.10. Then, 8,s

were chosen depending upon the levels of importance, inspectability, and

repairability of the design details.

It should be realized that in the foregoing design applications, the

design reliability (Safety Index, Prior probability of failure, PfD ) has

been targeted to a service life, Ts. For the lognormal formulation (Eq.

9.17), the safety index, P,, for any exposure period, (t) can be
expressed in terms of the design safety index, PD for a service life as:

In(t/T.) (9.40)
B, = flD- ~,nT

For t c Ts, the safety index is much larger than B, (Figure 9.11). This

explains why there is a very low probability of finding fatigue failures

early in the life of a structure (if all has gone well).

This equation also points out how inspections and repairs might be

utilized to maintain the safety index above some value (Figure 9.12).
Inspections can be used to reduce the uncertainties that contribute to

a,,,(Eq. 9.17), and thus increase ~,. Repairs (if effective and well
done) can increase B, by erasing all or a large portion of the cyclic
damage.

The optimum inspection and repair strategy will be a function of the

elements (details) importance to the capacity and serviceability of the

system, inspectability, repairability, and costs. The reader is referred

to reference [910] for additional details on the roles of quality

assurance and inspections in maintaining fatigue reliability.

246
 Desfgn Lffe, TD, fora Service Life, TS, of 20 Yrs.

400 200 50

*
5 - \ u

v
3 c
u

.
,-

2
.

.02 05 0.10 O*2O 0.50 1.0

Target Damage Level for Component, do .

FIG. 9.10 EXAMPLE: THE TARGET SAFETY INDEX

AS A FUNCTION OF THE TARGET DAMAGE
LEVEL FOR A COMPONENT

247
 [,
.10 g
I I I I I I I I I I
_ 10-8
,,.

\ 10-7
- *

\ 8 (pl.o _ 10+

[
\ _ 10-6
- Y,*
_l o 4
+.+g

.,~_l 0-3

\ . . .

1 ~,= 2.5> 1 0-?

I I I 1 I I I I I
o 0.2 0.4 0.6 0.8 1.0

ma
EXPOSURE PERIOD/SERVICE LIFE

FIG. 9.11

248
6
I I I
EPAIR +1\ REPAIR +2-

4
,0-4

3
/?-*--:, ----

10-3

[ REPAIR
THRESHOLD
L
MINIMUM
$
2 ,0-2

1 I I I 1 1 I I I I
0246 a 10 12 14 16 18 so

TIME-YEARS

F~G 9.12

24!3
26/
9.5 FATIGUE RELIABILITY - SYST~

Fatigue failure of a detail in a fail-safe engineered structure does

not constitute failure of the system (recalling the Introduction to this

Section). Thus, in addition to design of details for fatigue,

consideration must be given to fatigue reliability of the structural

system. The reader is referred to Section VIII for a discussion of

system reliability.

For a system of N identical and independent elements in series (a chain

system), the probability of failure of the system, Pfs, is related to the

element probability of failure, Pfe as:

P fs [l-(l-Pf,)~] (9.41)
=

or approximately,

(9.42)

If because of materials, construction, design, or loading, there is a

very high degree of correlation of the strengths of the elements, then

for perfectly correlated (dependent) elements:

P fs = Pf, (9.43)
...

Thus in a series system,correlationin element strengths has the effect

of reducing the probability of-failure.

250

242
Correlation expresses the degree of dependence (or independence) between

random variables (refer to Section II for a discussion of correlation).

Zero correlation implies independence. Unity correlation implies perfect

dependence.

Correlation is generally expressed by a correlation coefficient, dJV:

Cov(u,v) (9.44)
puv =
Uu,uv

where U and V are two random variables and the os are the standard

deviations of these variables. The covariance ofU and V, COV(U, V), is:

Cov(u,v)=q(u-pu)(v-pv)]

where the ps are the expected values, E[F],of the variables. The two

random variables are said to be uncorrelated if P = O (independent), and

to be perfectly correlated ifp = tl (dependent).

Cornell [9.11] has suggested a useful approximation for the correlation

coefficient between the probabilities of failure of the systems

components as:

v: (9.45)
P =
v: + v:

where VS2 and VR2 are the squared coefficients of variation of the

load(s) and resistance (R), respectively.

If the resistance is the dominant uncertainty, then P tends toward unity

(dependence).

251
Figure 9.13 [9.12] shows the Safety Indices for series systems, P,, as a

function of the correlation coefficient, the number of elements (N), and

the element safety indexp, (assuming normal distributed strengths and

equally correlated elements). For the high reliability elements

(B. = 3), the system safety index is approximately equal to the value
based on zero correlation (Eqs. 9.41 and 9.42) forP< 0.8. For high

degrees of correlation (P > 0.8), there is a small correction to the

element safety index to determine the system safety index.

For the fatigue reliability of a series system, the design probability of

failure of the N elements which compose the system (pfeD) can be

reasonably and realistically related to the system design probability of

failure (PfsD) as:

PfsD (9.46)
P,,~ = ~

Such an approach has been used in developing fatigue design criteria for

the connector elements of a tendon system for a Tension Leg Platform

(TLP) [9.13].

In the case of a parallel member system, the problem is much more

complex. Martingale and Wirsching [9.14] and Stahl and Geyer [9.15] have

studied the progressive fatigue characteristics of such systems. Typical

results for a system comprising parallel brittle members (B. = 3.0)

(2 joints per member) and a correlation in element fatigue lives of

32 percent is shown in Figure 9.14. The three curves are for:

252
 Ig
~N+

.
-------
O.g
z
Ulo
=U
ul\
. -------
0.6

/ N=2
/#~ N=5
/
/ N=1O
./ +
0.4

LLiJJ
O* ,.

Ow &= 3.0
i= 0.2 * .
.<
a ~e= 2.0 ---

0.0 I 1 I 1 i 1 1 I I

0.2 0.4 0.6 0.8

ELEMENT CORRELATION COEFFICIENT, p

FIG. q. 13 SAFETY INDICES FOR SERIES SYSTEM OF N ELEniENTS

253
 4

3*: 1

3 -A - - .. .A.

2.5

1.5
1 2 3 4 5 7 8

tWJMBER OF PARALLEL MEMBERS (2 JOINTS PER MEMBER)

FIG. ?. 14 SAFETY INDEX VERSUS NUMBER OF MEMBERS,

COMPARISON TO PROGRESSIVE
COLLAPSE MODEL &s3.0, p =32%

254
 a. First member failure

b. 50 percent member failure (50 percent loss in capacity)

c. 100 percent member failure (100 percent loss in capacity).

Only in the case of >50 percent member failure is there a beneficial

influence from adding parallel elements.

The effect of adding joints to the parallel elements for the condition of

100 percent member failure is sunnnarizedin Figure ~.15 (based on results

from [9.15]). The ratio of system to element safety indices (B,iB.) is

for correlations of 30 to 60 percent, 1 to 4 parallel elements, and 2 to

16 joints per parallel members. Adding joints to the parallel members

swamps out much of the beneficial effects of redundancy, and the system

behaves much more 1ike a series system (Figure 9.15). For example, for
an element correlation of 50 percent, Figure 9.15 indicates that for a

system of 10 elements (joints) in series, the B,tp. ratio is 0.5 and 0.75

forgo = 2 and 3, respectively. Referring to Figure 9.13 for the same

number of joints per member, BS/F, = 0.5 to 0.6 and 0.75 to 0.80 for

B, = 2 and 3, respectively.

Given a target reliability for design of the system (B,.), Figures S.13

and 3.15could be used to define the target reliability for the

individual elements (E.D).

Reliability based methods for analyses of complex structural systems are

being developed [9.11, 9.16], and these methods are being extended to

considerations of inspections and repairs [9.16, 9.17, 9.18]. The reader

is referred to the cited references for these research developments.

255
 P=30 -60%

N =1-4
PARALLEL
100% FAlLED
1.0

I
0.2

Q*O I I I I I I
2 4 6 8 10 12 14 16

NUMBER OF JOINTS PER MEMBER (N)

FIG. 9.1!5

256
 REFERENCES

9.1 Geyer, J. F. and Stahl, B., Simplified Fatigue Design Procedure

for Offshore Structures, Offshore Technology Conference, OTC
Paper 5331, Houston, Texas, May 5-8, 1986.

9.2 American Petroleum Institute, ReconmnendedPractice for

Planning, Designing and Constructing Fixed Offshore Platforms,
API Recommended Practice 2A (RP2A), Sixteenth Edition, April 2,
1986.

9.3 Wirsching, P. H., Probability-Based Fatigue Design Criteria for

Offshore Structures, API-PRAC Project #81-15, Prepared for
American Petroleum Insttiute, Dallas, Texas, January 1983.

9*4 Wirsching, P. H., Fatigue Reliability for Offshore Structures,

ASCE Journal of Structural Engineering, Paper No. 19235, Vol.
110, No. 10, October 1984.

9.5 Munse, W. H., Fatigue Criteria for Ship Structure Details,

Extreme Loads Response Symposium, Arlington, VA, October 19-20,
1981.

9.6 Wirsching, P. H. and Chew, Y.-N., Considerations fo

Probability-Based Fatigue Design for Marine Structures, Marine
Structural Reliability Symposium, Sheraton National Hotel,
Arlington, Virginia, October 5-6, 1987.

9.7 Munse, W. H. et al., Fatigue Characterization of Fabricated

Ship Details for Design, SSC-318, U.S. Coast Guard, Aug. 1982.

9.8 Marshall, P. W, and Luyties, W. H., Allowable Stress for

Fatigue Design, BOSS 82, Boston, Mass., August 1982.

9.9 Wirsching, P. H. and Light, M. C., Fatigue Underside Band

Random Stress, ASCE Journal of the Structural Division, Vol.
106, No. ST7, July, 1980.

9.10 Committee on Fatigue and Fracture Reliability of the Committee

on Structural Safety and Reliability of the Structural Division,
Fatigue Reliability: Quality Assurance and Maintainability,
Journal of the Structural Division, ASCE, Vol. 108, No. STl,
January, 1382.

257
~,11 Nordal, H., Cornell, C. A., and Karamchandani, A., A Structural
System Reliability Care Study of an Eight-leg Steel Jacket
Offshore Production Platform, Marine Structural Reliability
Symposium, Society of Naval Architects and Marine Engineers,
Arlington, Virginia, October 5-6, 1987.

9.12 Grjgoriu, Mircea and Turkstra, Carl, Safety of Structural

Systems with Correlated Resistances, Applied Mathematical
Modelling, Vol. 3, April 1979.

3.13 Chen, Y. N., and Wirsching, P. H., Probability Based Fatigue

Design Criteria for.TLP Tendons, Proceedings of the Fifth
International Offshore Mechanics and Arctic Engineering (OMAE)
Symposium, The American Society of Mechanical Engineers, Vol. 1,
Tokyo, Japan, April 13-18, 1986.

9.14 Martingale, Scott G. and~irsching, Paul H., Reliability-Based

Progressive Fatigue Collapse, Journal of Structural
Engineering, Vol. 109, No. 8, August, 1983.

9.15 Stahl, Bernhard, and Geyer, John F., Fatigue Reliability of

Parallel Member Systems, Journal of Structural Engineering, Vol.
110, No. 10, October, 1984.

9.16 Paliou, C., Shinozuka, and Chen, Y. N.,Reliability and

Durability of Marine Structures, Journal of Structural
Engineering, Vol. 113, No. 6, June, 1987.

9.17 Carr. P., Claytcin,?4.,Busby, P. L, and Dobson, J., A

Probabilistic Strategy for Subsea Inspection of Steel
Structures,, Proceedings of Society of Petroleum Engineers
European Petroleum Conference, SPE 15868, London, 20-22 October,
1986.

9.18 Madsen, H. O., $kjong, R. K., Tallin, A. G., and Kirkemo, F.,
Probabilistic Fatigue Crack Growth Analysis of Offshore
Structures, with Reliability Updating Through Inspections,
Proceedings of the Marine Structural Reliability Symposium,
Society of Naval Architects and Marine Engineers, Arlington,
Virginia, October 5-6, 1987.

258

.qo
 19. APPLICATIONS TO SHIPS AND MARINE STRUCTURES

10.1 Long- and Short-Term Procedures:

Two types of analyses may be used for assessing a vessel

strength under extreme load; short-term and long-term analyses. At
the design stage, if one knows the route of the ship and and if that
route is more or less permanent, then the probability of failure can
be predicted using long-term analysis. If on the other hand the
route of the ship is not defined, then the short-term analysis can be
used to obtain the probability of failure under one or more
conditions that are- considered to be the severest the ship may
encounter during its lifetime. An example of this situation are the
design conditions checked in the ASR catamaran by Lankford [10.1]:

(a) One year of continuous service in average North Atlantic

weather.
(b) Six months at a position in the North Atlantic where the
worst weather for this period of time would normally be
expected.
(c) Two months on station in the worst area and the worst
season in the North Atlantic.
,

A more simple short-term analysis criterion is to consider the -

single most severe sea condition (or a sea condition with a specified
return period) and subject the vessel to this condition for a specified
period of time.

These two methods, short- and long-term analyses, will

naturally produce different final results for the safety margins and
therefore care must be taken when comparing safety margins of
different ships, i.e., the method and criterion used in predicting the
loads acting on the ship will have a considerable impact on the
resulting safety index.

259

27J
 To further amplify on this point, the long-term distribution of
the wave loads acting on a ship may be determined by tracing the
expected route of the ship during its lifetime. Based on ocean wave
statistics along that route, the long-term wave load probability
distribution (usually taken as Weibull or exponential distribution) for
the entire history may be determined. Any lack or deficiency of data
on wave statistics over a period of time covering the ship life should
be corrected for. In the short-term analysis, a distribution of the
extreme load is predicted on the basis of criteria such as one
occumence in a lifetime, one extreme sea storm of a spectilc duration
or a short-term operation in a specific location under severe sea
conditions. For that purpose, one of the extreme wave load
distributions discussed under Prediction of Extreme Wave Loads in
Chapter II is used.

It should be noted that there is a fundamental

difference
between computed results based on these two avenues. In the
short-term analysis, the computed probabilities of failure are
conditional probabilities given the occurrence of an extreme wave
load per a selected criterion, Care must be taken in this case in
determining the response of the ship to this extreme load since non-
linearities may play an important role. In the long-term analysis,
however, the resulting probabilities of failure are associated with the
entire history of the expected load acting on a ship during its
lifetime.

A. Procedure for long-term analysis

The following procedure may be used for calculating the

probability of failure during the ships operational lifetime.

1. Define the mission profile for the ship that includes

estimates of
a. ship route
b. expected total years of service

260

272
 c* Number of days per year the ship is expected to be
at port and underway
d. Nominal cruising speed and maximum speed and
the corresponding fraction of time during operation.

2. From the ship route and available wave statistics such as

in reference ~Q2], obtain the frequency of occurrence of
different sea conditions the ship will encounter in each of the
geographic areasl 1.

3. From step 2 above and the mission profile of the ship

(more specifically from expected number of days in each
geographic area), determine the total frequency of
encountering different sea conditions.

4. Determine the root-mean-square value K (rms) of the

wave bending moment in each sea condition. First, the
response amplitude operator, RAO, has to be determined either
from available strip-theory (seakeeping) program or from
model experiment. The rms values can then be obtained using
the determined RAO in conjunction with existing sea spectra
such as Pierson-Moskowitz ~0.3]. These programs usually give
the value of the stillwater bending moment also.

5. From the total frequency of encountering different sea

conditions (step 3) and the rms values of the wave bending
moment in each sea condition, determine the average wave
bending moment l,.

11 In almost all the main areas where ships operate, statistical data
concerning wave heights and periods have been observed and
tabulated. The surface of the earth is divided into a grid of ten-
degree squares known as Marsden squares. These squares are
arranged into geographic areas over which wave conditions are fairly
uniform. The areas are given a code numbe~ see, for example
reference ~0,2].

261
 6. Estimate the strength parameters (mean ~ and standard
deviation u or the coefficient of variation) including the
objective and subjective uncertainties. Each failure mode will
have its corresponding y and 6.

7. Calculate the probabilities of failure in the different

modes 12 (that is, yielding or buckling at different locations) in
sagging and in hogging. Combine to get the total probability of
failure.

A block diagram of the foregoing procedure is given in Figure 9.1.

B. Procedure for short-term analysis

The following procedure may be applied in the short-term

analysis.

1. From the assumed design criteria artd ocean wave

statistics, calculate the frequency of operation in different sea
conditions.

2. Calculate the rms value of the wave bending moment in

each sea condition using either strip-theory approach @4, 10 .7]
or towing tank experiment results in conjunction with available
sea spectra. Calculate also the stillwater bendirig moment.

3. Estimate the strength parameters (IL and 6) for each

failure mode.

12 The probability of failure in the different modes can be calculated

using equation (4.35) if the stillwater bending moment is considered
deterministic or from equation (4.38 or 4.41), if it is considered as a
random variable. In the former case, if desired, different values of
the stillwater bending moment could be used with an estimate of the
corresponding fraction of time.

262
 4. Calculate the probabilities of failure in each sea condition,
i.e.,
P [ R ~ lfith sea condition ]

5. The total probability of failure is thus

p* P [ R 5 1 / ~th sea condition ] p(l)

z
i
where p(i) = probability of operation in the ith sea condition as
determined from step 1.

A block diagram of the above procedure is given in Figure IQ2.

If only one storm condition (with a certain return period) is

specified in the short-term analysis, then only the probability of
failure in that condition needs to be calculated. This probability is
thus a conditional probability given that the vessel encounters the
specified sea condition, and, is expected to be larger than the long-
tenn or life probability of failure.

In general, long-term analysis requires much more information

and computational effort than the short-term analysis. Long-term
analysis is, however, necessary if fatigue failure is considered since
the entire history of loading should be included. On the other hand,
failure under an extreme load can be more easily estimated using
short-term analysis.

10.2 Application Examples:

10.2. I, Short-term analysis - Level 3 Reliability

In this example, we will evaluate the probability of exceeding

any limit state of a ship during a specified storm condition. The limit
state can be an initial yield limit state, an initial buckling limit state,
the ultimate strength limit state, or any other condition desired to be

263
evaluated. The duration of the presence of the ship in the storm is
limited by the stationarity condition to a short period of time.

Consider now a tanker of length = 763 ft, breadth =125 ft, and
depth = 54.5 ft. We will evaluate the probabilities of exceeding the
wa4Am
..t1b(ww
d4y*
Wd9m.y
, irramant 904 -~,im,
U aOErm,WAq
mAMl
-lm .X)
om, lrwnw d UWr9a I.
- mrawu *ZM
/ )
J
wti~e
--
Aiimmws-=w- 1

1
1A J
still
tmndlw
W.t.r
-*
L [ ,.
ltr*wth
?*b--f,

Figure 10.1.
Iz!@izz
Long-Term Procedure

.E!e
Tha R.M. S. V41U90 Of
wave bending -nt in
diffmrorrt soa conditions
fromstrip theory or X.1 test

I still Watwr
Bondinq mnne
conditional
f ilura in ach

,..
probability of
e= condition
-l

1
---4 Total probability
of failure

Figure 10.2. Short-Term Procedure.

26A
initial yield limit state and the ultimate strength limit state in a
storm condition specified by a significant wave height of 29.0 ft and
an average wave period of 10.1 sec. The storm is assumed to be
stationary under these conditions for a period of one hour. The
following parameters were computed for the tanker using a typical
ship motion program:

Stillwater moment (full load) ms = 669,037 ft-ton

(considered deterministic)
Rms of the wave moment m. = 216,450 ft-ton
r
Average wave moment period = 12.1 Sec
Bandwidth parameter of wave moment spectral density e = 0.337

Number of moment peaks (in one hour) N = ~ = 297.5

12.1

The mean and standard deviation of the initial yield limit state
were computed to be P.r = 2,420,488 ft-ton ~d ~r = 314,663 ft-ton,
respectively. The corresponding values for the ultimate limit state
were computed to be ~r = 2,804,760 ft-ton and ~r = 392,666 ft-ton.
The mean of the initial yield limit state was simply computed as the
minimum section modulus of the hull amidship multiplied by the
average yield stress. The mean of the ultimate strength was
computed using USAS, an elaborate nonlinear finite-element
program (see 605]).

Using order statistics and the determined mo, E, andvalues of

r
N, the expected maximum wave bending moment in N peaks is
computed from equation (2.57), to be 763,859 ft-ton. If one assumes
an ideal narrow-band case, such as e = O instead of 0.337, one
obtains, using equation (2;57) the slightly more conservative
again,
value of 767,543 ft-ton. That is, the error due to the assumption of G
= O is less than 0.5 percent.

Equation (2.58) may be used compute the extreme wave

to
bending moment with probability of exceedence a. For example, the
value of the extreme wave moment with a probability of exceedence
a = 0.0001 in the given storm condition is computed to be 1,086,685
ft-ton.

A semiprobabilistic factor of safety may be defined as the

resistance mean ~r divided by the sum of the stillwater bending
moment and the maximum expected wave bending moment as given
by (2.57). The computed values of this factor of safety in the given
storm condition are 1.69 and 1.96 with respect to the initial yield
and ultimate strength, respectively.

Finally, the probability of, exceeding a limit state pf, which

combines all the given information on the ship and the storm
condition, is computed using the basic reliability equation for a
normally distributed strength (see equation (4.5)).

pf=l. -+[2; 12 dz@,, .

I
. k -Fz~(z) e

where FZN (z) is the extreme total bending moment (stillwater plus
extreme wave bending moment). Several distributions can be used
for FZN as discussed earlier in the section entitled Prediction of
Extreme Wave Load in Chapter II. We will use FZ ~ (z) as predicted
from order statistics given by equation (2.54) with Rice distribution
given by equation (2.38) as the initial distribution. In the latter
equation, ms , the stillwater moment is added . Substitution of these
equations in (10.1) and carrying out the integration numerically we
obtain the following values for the initial yield and ultimate strength
limit states:

pf (initial yield) = 1.19 x 10-3

pf (ultimate strength) = 3.13 x 10-4

In order to examine the effect of the sea state on these

probabilities, a storm condition characterized by significant wave
height of 38.75 ft. and average wave period of 11.5 sec is considered

266
next. The ship motion program computed values for the rms wave
moment ~ , the bandwidth parameter e, and the average period of
5
the wave moment are, respectively, 2.863 X 10 ft-tons, 0.364, and
13.0 sec. It should be noted that the rms value of the moment is on
the conservative side because of the linearity assumption. The
number of wave moment peaks in cme hour is thus 276.9. Based on
these values and the resistance parameters given previously, the
following probabilities of exceeding the limit states where
determined:

pf (initial yield) = 1.23 X 10-2

Pf (ultimate strength) = 2.73 X 10-3

It is interesting to note that when pf for the ultimate strength

case was recomputed using equation (2.60), which is based on
asymptotic distribution instead of order statistics distribution, a
value of pf = 3.14 X 10-3 was determined (compared with 2.73 X
~()3 based on order statistics distribution), As expected, equation
(2.60) gives an upper bound on pf and its accuracy should increase
as N+Q.

It should be noted that the computed probabilities given in this

example are conditional probabilities given that the ship encounters
a specified storm for a specified length of time. They are
fundamentally different from those calculated by constructing the
long-term distribution of the wave moment along the ship route
during its lifetime (see LQ6]). The elaborate procedure in this latter
case produced unconditional lifetime probabilities of failure.

The main advantage, however, of the presented storm-based

procedure is its simplicity and consistency. From the environmental
data along the ship route (or structure location), a design storm
condition can be postulated and the probabilities of exceedence can
be immediately determined from the simple results given
previously.
267
 The procedure can be used to determine the average
probability of failure (or exceedence of a limit state) during the
entire duration of a storm rather than just the severest one-hour
period of the storm. In this case, a simulation of the storm condition
during successive short intervals of time (say one hour each) is
necessary. During each interval, the waves are assumed to be
stationary and may be represented by a pair of significant wave
heights and average wave periods. The rms values @ of the wave
bending moment can be calculated for each pair and the
corresponding probabilities pf are determined from (lQl). The
average probability of exceeding a limit state during the entire storm
duration is then

pf = 5P~Wfi;~fi=l
i=l i=l

where fi is the frequency of occurrence of the ith pair of significant

wave heights and average wave periods, and n is the number of
stationary short intervals during the storm.

The important high-frequency moments which may increase

the bandwidth parameter are due to either springing or slamming.
It is unlikely that springing moment is of any appreciable value in
high sea states where wave periods are typically large. Therefore
any increase in the wave moment rms value ~ will be negligibly
small. Slamming, however, may have some effect on the rms value
of the wave moment for small ships. It may be combined with the
wave moment to obtain a total rms value using, for example, a
procedure developed in ~0.8]. It should be noted, however, that the
underlying combined process of the wave and slamming moments is
not ngeneral Gaussian except in one limiting case when slamming
decay rates are negligible in comparison with the mean slamming
rate.

26!3

280
 Equation (IO.1) which gives the probability of failure has been
plotted as a function of non-dimensional variables for the case of
FZ ~ (z) estimated from order statistics with Raleigh distribution as the
initial distribution (i.e., Rice distribution with & = O). This
approximation leads to conservative estimates of the probability of
failure as discussed earlier. Figure 103,10.4,10.5,10.6 and 10.7 show the
value of pf as a function of G* = Q , N and v*= ~r - ms .
Fo no
r
As an example of the approximation involved, the probability of
failure computed from these figures for the initial yield limit state
and the second storm condition (significant wave height 38.75 ft.) of
the above example (with e = O) is 1.40x 10-2 compared to pf = 1.23 x
1()-2 obtained earlier. These figures thus will give slightly more
conservative values for pf but eliminate the necessity of numerical
integration.

.-

, .

26$1
 10 :. T,,,,.,.,
Tr
:;:j
...,:..
;
i:~:~..;...
I ,.
.:
,.:.::
:.::
.,,,
.,..
. .,

J
1 .,,,..

.i
. . . . . .. . . ,, I.-i...
I.J: ,.,4 . 1, . . . . . .

- . -,.. .
:.. . ,.. .
II
;:::: :::
:.:::::: ::: . . . . .
. . . . . . . . .: .. -. . .

.;. . ,..
.... .
. . .. . . . . . . .. . . . ..-.

. . . . . . . . ,. . . . . .

,.-. ,.
II
.,1. . . . ..4 .

!- .-+.- ------ .. ,. . . . -.

-.,. - :::: . :.:: -. . .

,.. . ,....- .,.,: .,. . . ----

.,...,=
+
4
. .
. .
.
.
. ----- . . ..-
----- ., ---

,
,..

.-. . ------ -.. .

...-
----

.,--- ----
. . ...

,.
4--.-
,-.-
. . .-

::. : -. =-=
,.-. .
!.. _
,..
. .
.
. -
I&
o .-- --

,,-
. . .

.+. - .
,+_ ~ .. -
,+- --dT- -
\ 1

A!&
,
,:
,1;

:= - ==7
--
=
-- :-. _ =
.
-- c
..
-
--
-.
:! -.--i- .T ;, -

S
,,, , :1
--if+ ~
:.) ,:,: 1,: !::!: .!
-
~
t-+ ~ ,, ill
10-7.
:1 :I ~! ~:~ :: I % 1,
.--+ .- n ==7
-
-
--

= ..
- $
: - --
---
..
~= . ...
.
-.
---
. ---- . .. . ~ - .+.
J
.- -
---

=F+l=l= ---
- . .. _..
: ---
,,
4+--
.--.]LJ. -..

2242 ,

m
1o-8- --.. . - ..-
~ ... :, ::
_. --.,. -,.
~ -- .+-- .- . 1 -+-- .,.
+ +-, .. .+ .- 1- L_
7-A
*.. . ., .
--- -4 .- .-*. u___ .4+ ,.- ,. A . ..- ,,

-.4- J .~. *
1= .-4 4+ .. :-.
:, ~:~
-- .: .:+ L+ :: . ... ~; :,4. :;

J
; p:
!;

1.-9
.!,
.,
-
:,;1
7 1

Figure 10.3. Probability of Failure.

270
 - - -
...
. .
::!:

.
.
.
T,,
:,,
. .
:.. .
.

..4. ;., .
. . . . ,.. .

.,..::: :.:: -
::.
.
:
. . . ,. ,., :.. :
.
... .. .,
. . .

... ,. .,: .
,..
.-
,.. . ..

-..
.
.
. .

.
,.,

.-
----

..
:.. ,
-.-.
....-
.
..-

:,
..
---
. .,. , ::
.. . .
,:

::
..
. ..

---
... ,

d
-..
-.. N. 539:.
:;:,:
2

.+.
. .. .40~ .._ ---
0--- . 300
~ 250
-___ =
----.
. .
-
. .._ . -. . . .
--
. -:. ----
----
. ~ ; >0---: ---;-

!.
~

~
=

7




--i-
II 1
! 4
;!:
. =
. ..-+...
.-........ .-----
.+-------._ +... =
. .. . -+. .. ... .. .. . .-
.+- ..... ... . ... . . .. -----
. . . . . ..+_+ .
-
. ----
.. . . ... . . . . .._. . .._ . . . . .-. ,
,.-. . ..+. ... . ,,. ..-< _._j

,.:: -.+,
:.: : .. . .
---
-4. . . .
T
.
; ,.
:,
.+ +:-ii.::::
::;+:
:;::
u *R
..+

-.+-.... .. ;~~jf:::.::
.-. :--
. .... ....
,,
.,-..,.. ,, ...._,., ,,., ,,,
,,
........ ...+.,, ., ....
10
5 9 )

Figure 10.4. Probability of Failure.

271

-w
 10- .,.,::,
::: ::; T
-
..::

T ,,. ....
.!, ,

:.. l! .,,
,.,4
.. . .
,1 . ...! ... ,..
!.. +
.
. ..-
., ., . . . . . .,.4 ... .
,.1 ,.. : 1 ..-

1 :14,..
I
.,4. l.. . . . . .
,.
,!

,, ,,

-
::: :.:::
.,,
- - -

. . . . .. ..
,,. . ,.. .
. .
..4
,.. . .,. .
::: ,.. - ,.-+
;+ -.. .
.,.

. ,.+4 ,.,

., !.,
-.. ,

.- .,. ,..

. . . ,.. -

.
:x ;:
. . . .

10-L ,. .,.
I I
.
I
. . .
1> d ,.. .
s+ .-.- WJ . . . ,..
-..
:::.
--.
- ..

x .- .
. . .. . . .. . . . ... . . . + !. :g
,.. ..- -.

+ Ilv ..- . ,

..m . .-, - .- .

!!i----
a
,... ...
......_
.
i --..-- .------ ---
- --+ ---

,-
---1. - -.

----
-. --

=:
IL
0 .+- .
--- - -

-

1
=
Y
1-4 =&
?- A
~_. = =

.
L ---
!,
--
.
... .

,-.:
, , .
,,
+ --
I ,, y ~ ,1, : l., !,
/1 ~

1;:1 11; Ill ::: -T

.1 :il :1
,!1 ;1 1! 1, :1
.4 i : i,
107. - - ~. - -. -. -- ----- = = =
. .- ----- .+. .
:_ .- =. . . -- .+-. ., . -+ .
. . .._- -- . . -, .
--- -- . --- : --

.-.. ---- ---- :_- -,---- __l. . - +- . _.+-

1
-.:- .-.. _.. . _ --- . . . -J ,i. .A, .. ---- .
T,,
. . . .-+. :---- _.. . _., . . . . . .. -+- -- -----

)
?1 : ,,
,,
10-8 . . . . +. .. -.. . . . . . . . . . .. .-_. - --..
.+.
-.
. .
I .-.

,..
.
-. ---
---- ---- -AL ----- . . . . . . .- .+.

,.
- ...: .,*=
.++.
!_. .

-+,--- .,,
. ..-.

-:--

-.. .
...

. ..
-

. . .
. . .

..- ~:y. . . . . . . . ...<

1 --,4, ,.. .

>
. . J.:.

,I

6
.+

::
.,,, .!--

7
. . . .

8
. ,!

1
9
.,. .

10
L 11

Figure 10.5. Probability of Failure.

272
11,- -

.
::,
,,, ,,
!,.4
. .
.
.
.
.
.
;.
,,
........
..........
.11 ..4. . . . . .

... : ..:4 ,.. 1 :. .,. . .

. ... ~
,,
,.. . . . . . .,. . . . . . . ,.
I
,
,,
- -
::. :
.... ::::. .:;
.1-. ;:1; . .
,,.
::.1..: . . . . . ---
.. .
..-. . . .
. ..-
. ..-

. . . . . ..- -. .A4- . ---- . . . . .

... . . . . !---- . ++-
. . . . . .. .
.+. . .+74. . . . . . .
.+ . ..-
. . . . . . . .

10-4
::::
-. .. .. . ........
----- I
::;:
. . . . . .
. ... . ::::.
..
. ..... .. .. . . .: . --

:;l-
.,. .
.4. .
-----

.--.. . ... . .
----- . . . . . ..-
. . . .+.+-.. .
---- .._....-+ -.

. ..-. .. ----
tt
.-.
-
-.-4 ----
---1
:--1
,.

-.
. -
- 4--
-------

.
---

. ..-.

.
.- ..
- . .
x:
--
..
*

-:

?J$ ... ,
4 - .+ 7-
l-+ -. .
.,,
J-

=
-
,. 2%:
4= ---- e
%<

-
.-,,
-. L-

~-q7 -

y-

-,
-
.. .,
--.. + v=
-- - . .+
.
.

3....,. --- --_T

-. . .-+ .... .-- . . . . . . . . . . . . ..- . . .. .

;___
..+.-
--- ~
-..++ Zt=
. ..
..
.
-..
+ -4..
.;

-
:::: . ,..-.
.. .. :: :: ...
----
---
. . .. .. . :::=. .v
.. . . . . . . . .---,
. . . . .- , ..---- ---, .
.. J- .-. .. . .... .... ...... --l
.
_,:i
1*2-.... .. ....
. ;.. ,
.... ...- .. .. .... .... .-.J
*
6 7 8 9 10 11

+ %-m,
P= Jq

Figure 10.6. Probability of Failure.

273
 ,, ,. ,:
:
,
.,,::,
,,,
,.. ,,
,..
..,:
1:1/ T,,.+.
,.,
.,
,.,
;.,
.,,
,.
T....,
.,.
,,,
j:

:..1
:::...

. .
;;1,

.
:..
,.,
,,,
.
-
,.:.
.
.,. . :.
,,
,.
,.
. . .

It
;,!!

.. . .,

-.. . .. ... . . . . . ... :. . ,.. :.. .

1~,, ,
-7
~~-,, l, i,
1 I 1 I
. . . . . . . . . . . . . . . . . -
. .. . . . . . . . .
.:: :::: ;:::. :::
. ... :
,., ,
.
. ..J ...
. . -. . A_- +.. . ..- J . , .,. .
. .,
. . . . . . ..-
. . . . . .. . .. . . . .,. . .-.I . ,.. .,J
,:
-4
10 / /
.-. , .,. .
i . . -

--_ -
::
-..
.-. .
..
----- . ..-. -.
w ,-. . ---
m . .-..

..-
,-- :
-
--
-.. .

.
---
.-.
d ,,
-

,,
----
u.

--- ---
,.,
-+. .

n
,

m
-+...
-.
o
Z ---
q!
-!19 .
.-. ~..- .
1, ----

Ia!iikE
,< ---- ,-
Eo

-.
_ .
.

-
:

-+
..! .- -

.-
- -..
-
&<
:Io-d : : ,
.+
-----

) : ;

la
T _
m
.-

,,
Zi.
a,, ,, 1<1

---
..
-
>--

.-.
,
.
..
:...
+

------
---
....--, .-
--
I , ; :!, l!;: T ;,,
,1 1, : 11! - -1
,,,1 t

-M!#&!
,!:

--
+.++-.
(,

---- --------
--.+.
%- .-.
.. . .

1==1=+t=l= I
.L

m
. . . . . ..
. . .. .
. ..

&
!:
10-0 -.. ., ---- . . ... . . . .. . . . -,. .
..- . . ..- -- _,-. __ --, -::
-. ..-. . ----- . . . . . . -.
....-;:.
----- . . . .. . ..
_+_r,----
. .. .... ...44:-....+ -- +
....._ .,. ...
.

;..
;,,
..+.
...4....
. .

..-...+
.. . . . .
!-
. .
-:--
.-

J:.-.
.
.~-
....
.... ....,.:.. .....
..+4
.: -4.-
,+
L1.,A
J..
. -----
. .
-
.

10-9
5
.- a;: !? :
4...+

b
.0:
i
7
I I

6
l. +--.+,.

,.

9
!.: .-:,

10
.4..

11

Figure10.7. Probability of Failure.

274
10.2.2. -- Level 3 Reliability Based on Four Different Extreme
Distributions.

Short-term Analysis:

The impact cm the probability of failure of the different

extreme value distributions of wave bending moment discussed
earlier in Chapter 2 will be examined in this example. The
distribution function FZ ~ (z) in equation @l) will be substituted for.
using equations (2.56) 9 (2.60); (2.70) and (2.71) in order to
obtain pf based on order statistics, type I asymptotic distribution,
uncrossing analysis and a two state description of the random
process. The tanker cited in example 1 is used again with bending
moment parameters (the second storm condition) given by:

ms = 669,037 ft-tons ; m. = 286,300 ft-tons

{

N= 276.9 . = 0.35
* q

The uItimate limit state was considered (pr = 2,804,760 ft-tons

and ur = 392,666 ft-tons). The results of the probability of failure
are shown in figure10.$ and table 10.1. As expected the probability of
failure based on the- asymptotic distribution is higher than the rest
and, in general, the ,agreement between the other three distributions
is very good.

275
 4.00 .
. _________ .______ -- __ . .

1
m
o
-3.00 .. _

.
x

C -t
&
+-
0
r /
/ .

n
n
0 -1
+rnmm
& o.o~O.do mlrlnn nT-n-l-?-r ~
50. 1m 0 K) 150 )0 200:00 25C )0 300. 00

Figure 10.8 Probability of Failure According

to Four Differetqt Extreme
Distributions
 TABLE j~.~ PROBABILITY OF FAILURE ACCORDING To
FOUR DISTRIBUTIONS
Two-State
~ Order Statistics Asymptotic Dist. Up-crossing Description
?ype 1 Analysis (q*.35)
10 0.2&+183x 103 0.711313X 1O-J 0.280864x 10-3 0.268484X 10-3
20 0.486767X 10-3 0.856830X 10-3 0.4G3547X 10-3 0.463766X 103
30 o.f555854x103 O.lolj?sx 10-2 0.652819x 10-3 0.627937X 10-3
~
0.804382X 10-3 0.116312x 10-2 0.801514X 10-3 0.772696x 10-3
50 0,928461 X 10-3 0.129867x 10-2 0.935738x 1o-3 0.903715x 10-3
60 0.106161 x 10-2 0.l&2424X 10-2 0.105901 x 10-2 0.1O2426X 10-2
70 o.l17608x 10-2 0.1*149X 102 0.117359X 10-2 0.113652 X 10-2
80 0.128~ X 10-2 0.165174x 10-2 0.128105 X 10-2 0.124191 X 10-2

90 0.13&83 X 10-2 0.175600x 10-2 0.138252x 10-2 O.1%1* x 10-2

100 O.1481O9X 10-2 0.185508x 10-2 0.147887X 10-2 0.143621 X 10-2
110 0.157291x 10-2 0.194963x 10-2 0.157075 x 10-2 o.152656x 10-2
120 0.166080X 10-2 o.204017x 10-2 0.165871 X 102 0.161312 X 10-2
130 0.174522x 10-2 0.212713x 10-2 0.174318x 10-2 0.169629 X 10-2
140 0.182651x 10-2 O.221O87X 10-2 0.182452x 10-2 o.177&t2x 10-2
150 0.190496x 10-2 0.229169x 102 0.190303X 10-2 0.1853W X 10-2
160 0.198085X 10-2 0.236985x 10-2 0.197897x 10-2 G.192867X 10-2
170 0.205438X 10-2 0.244558x 10-2 o.20525+x 10-2 0.200124X 10-2
lea o.212576x 10-2 0.251906x 10-2 0.212395 X 10-2 0.2071?0 X 10-2
190 0.219513 X 10-2 0.259047X 10-2 0.219336x 10-2 0.21M21 x 10-2
200 0.226264 X 10-2 0.265996x 10-2 0.226092 x 10-2 0.220691 X 10-2
210 0.232W x 10-2 0.2Z766X 10-2 0.232674 X 10-2 0.227192 X 10-2
220 0.239262 X 10-2 0.279369x 10-2 0.239096 X 10-2 0.233536 X 10-2
230 0.245529X 10-2 0.285815x 10-2 0.245s66 X 10-2 0.239731X 10-2
240 0.251654x 10-2 0.292114x 102 0.251494 X 10-2 0.245788 X 10-2
10.2.3. -- comparison of Level 2 and Level 3 - ~ect of
Correlation Between Wave and Stillwater Bending
Moments.

Long-Term Analysis:

This example consists of two parts. In the first part we will

discuss a long-term procedure applied to the ship used in the
previous example. A comparison will be made between Level 3
(exact) and Level 2 (approx.) methods of reliability analysis. In the
second part of the example we will examine the effect of correlations
between the stillwater and wave bending moments on the
probability of failure using again a long-term analysis.

The ship is assumed to have the following mission profile,

a) ship life = 20 years

b) ship in port 65 days/yr. and underway 300 days/hr.
c) ship route: Marsden square numbers, 1, 2, 4, 12,
21, 23, 25, 30, 31 (see Figwe 5,3)
d) time proportions in Marsden squares: 2, 2, 1, 1, 1, 1, 1, 1, 1

The frequency of occurrence of different sea conditions

specified by the significant wave height is calculated in each
geographic area. For the ship, frequency of encountering the
different sea states during the operational lifetime is obtained using
such information in conjunction with (a), (b), (c) and (d), and is
shown in Table 10.2.

To calculate the number of wave bending moment (or, wave

peaks (N)) the ship will encounter throughout her life, at different
sea states, first we calculate the average wave period at different sea
states from wave data (Ref.l!l .2) as shown in Table 10.2. The ratio of
number of days the ship spends in a particular sea state and average
period of waves in that sea state gives the number of peaks. Such
results are also shown in the same table. ~

278
q o
 The root mean square of the WBM ( ~ ) in each sea state are
calculated from standard sea-keeping program and are shown in
Table10 .3. In the same table the scale and location parameters ~Ni
and UNi of the asymptotic distribution as calculated from equations
(2.65) and (2.64) are shown (G = O).

The SWBM is assumed deterministic and all variables are

assumed independent. The complete problem reduces to calculation
of P (or, pf) for the performance function

g(~ = R- Ms-Mw

R* Normal (2420488,3 14663) ft-ton (strength)

Ms M Deterministic = 669037 ft-ton (stillwater)
Mw N Extreme Value 1 ( up~,u~~(wave)

In the exact method the probabilities of failure p; for each sea

state i are calculated by numerical integration of the equation:

+tn
-UN-(m-M*-uN ) 1
P. l-
L
,lW
I
_@ Xp(-e
L i.-
T ( :
r
)) dm

The results are shown in Table 10.3. Similar calculations were

done by the advanced Level 2 method and are also shown in the
same table. The lifetime probability of failure (exceedence of the
initial yield limit state) can now be calculated from,
.
pf = Pi i
z @o.2)
i

where fi is the frequency of occurrence of such sea state. In our

application example, the results are,

pf exact = 5.2268 X 10-3 (Level 3)

p f approx. = 4.7851 x 10-3 (Level 2)

279
 In the second part of this example the SWBM is assumed to be
non-deterministic and correlated to the WBM. Only Level 2 method
w&s used to find ~ (or, pf), where,

RN Normal (2420488,3 14663) ft-ton

Ms ~ Normal (403520,161408) ft-ton
Mw w Extreme Value 1 ( ~Ni, UNi)

The correlation matrix is:

1 0

II
1
R=o 1
23
0 P 32 1

To show the effect of such correlation we calculated failure

probabilities for P ~1 = ~12 = 0.0, 01, 03, 0.5, 0.7, 0.9. The
results are presented in Table10.4 and Figure10 .9.

A comparison of results in Table 10.3 reveals that in our

example. the Level 2 method yielded lower values of failure
probabilities than those obtained from the exact method. The extent
and direction of difference in results between approximate and exact
method depends on the nature and shape of the nonlinear
transformed failure surface. As an example, let us consider the
failure surface in sea state 10 in the first part of our application
example. For the linear failure surface R-Ms - Mw = O in the original
space, one obtains, by transformation, the nonlinear failure surface in
the standard space as,
1
(Pr+YrUr)-M= - { U; & in in [ ]} =0
N *WW)

00.3)

280
where Yr and Ymw are standard uncomelated variables of R and Mw
respectively and are expressed as,
r-p
r
Y=
r u
r
@4)

- UM(mW - u,)
Y= ? [ exp [ -e )1
mw
@Q5)

Substituting the appropriate values in equation (lQ3), the

failure surface is obtained as shown in Figure 10.10. The area of the
single shaded region in this figure represents exact failure
probability. On the other hand, area of the double shaded zone on
the failure side of the linearizing tangent line represents the
approximate failure probability according to Level 2. Since the area
of the double shaded zone is less than that of the single shaded
region, the approximate method is seen to give lower values of p f.
Instead of being concave, if the failure surface is convex, such
linearization would have yielded higher values of pf.
.

While it is attractive to use approximate methods for ease i n

calculations, one rlmst have an idea of the failure surface for t h e
problem under consideration. The extent of approximation is well ~ ~
understood by having this surface drawn. If this surface is highly
nonlinear, first order approximation analysis may yield gross error.
However, for most practical cases, the problem is not very acute and
the approximate method of level 2 would suffice. Also, if the surface
is highly nonlinear, one may use a few linearization points and
express the actual failure probability in terms of bounds.

Results in Table 10.4 and Figure 10.9 show the effect of

correlation between Mw and Ms. As expected, as the extent of
positive correlation grows, failure probability, too, increases. From
zero correlation to almost full correlation ( P = 09), failure

281
probability increases by about 25%. Such an increase in the
probability of failure is not considered to be significant and, in fact,
in terms of ~, it would be very small. The results indicate, therefore,
that the correlation between the stillwater and wave bending
moments is not important and may be neglected in future analysis.

Sea S&g. Wave Sea State Number of Avg. Wave Number

State Height Frequency days Period of Peaks
(i) (ft. ) (Sees. ) (N)

0 7.15 0.0737246 537.40 5.721 8115952

1 7.80 0.1019174 743.00 5.897 10886078
2 9.15 0.1808501 1318.40 6.031 18887375
3 10.49 0.1592058 1160.60 6.555 15289703
4 11.84 0.0970026 707.20 6.987 8745109
5 13.18 0.0710068 517.60 7.253 6165813
6 14.53 0.0449062 327.40 7.484 3779711
7 15.88 0.0318023 231.80 7.641 2621060
8 17.22 0.0196168 143.00 7.783 1587459
9 18.56 0.0182419 133.00 7.847 1464407
10 19.90 0.0032569 23.80 7.831 262587
11 21.24 0.0035286 25.80 7.951 280357
12 22.58 0.0050191 36.60 8.067 391997
13 23.93 0.0040398 29.40 8.069 314883
14 25.27 0.0017686 12.80 8.023 137847
15 26.61 0.0019604 14.20 8.102 151429
16 27.95 0.0020007 14.60 8.239 153106
17 29.29 0.0008293 6.00 8.453 61125
18 30.63 0.0008039 5.80 8.287 60471
19 31.98 0.0015291 11.20 8.481 114100
91 36.00 0.0000232 1.65 8.771 16254
92 38.68 0.0000389 0.29 8.609 2910
94 44.05 0.0000059 0.05 7.420 582
95 46.73 0.0000149 0.11 8.273 1149
port 0.1780822 1300.00

Table 10.2. Estimation of Number of Wave Peaks the Ship Faces at

each Sea State, During Operational Lifetime.

282
 .
Sea R.M.S. of Scale LOcatlon Exact Failure -
Approx.
State HBl?( m ) Parameter Parameter Probability Failure
(i) (ft-ton) (qj J ( UN;) Probability
pi&xact
p-~lpprox
au 21650 2.6055e-04 1.2212e05 3.2187e-06 1.1515e-g7
01 26700 2.1321e-04 1.5199e05 3.2187e-06 1.9185e-07
02 40415 1.4323e-04 2.3395e05 3.2187e-06 7.4390e-07
03 55570 1.0351e-04 3.1964e05 3.5167e-06 2.8659e-06
04 76500 7.3909e-05 4.3253e05 1.6212e-05 1.5214e-05
05 100315 5.5743e-05 5.6095e05 9.2506e-05 8.7141e-05
06 129905 4.2367e-05 7.1495e05 6.1077e-04 5.7095e-04
07 155885 3.4877e-05 8.4751e05 2.5651e-03 2.3866e-03
08 180420 2.9618e-05 9.6411e05 7.8703e-03 7.3035e-03
09 19918S 2.6752e-05 1.0614e06 1.7956e-02 1.6695e-02
10 221560 2.2548e-05 1.1068e06 2.6627e-02 2.4546e-02
11 241045 2.0779e-05 1.2073e06 5.3888e-02 4.9997e-02
12 262695 1.9320e-05 1.3332et16 1.1368e-01 1.0663e-01
13 278570 1.8063e-05 1.4017eE!6 1.6216e-01 1.5281e-01
294450 1.6522e-05 1.4325e06 1.9008e-01 1.7899e-01
:: 308880 1.5813e-05 1.5086eB6 2.9189e-01 2.4856e-01
16 324040 1.5C80e-05 1.5834e06 3.4412e-01 3.2866e-01
17 338470 1.3811e-05 1.5891e06 3.5483e-01 3.3817e-01
18 350740 1.3379e-05 1.6459e06 4.2291e-01 4.0512e-01
19 363730 1.3268e-OS 1.7553e06 5.5559e-01 4.6154e-01
91 398370 1.1054e-OS 1.7543e06 5.6263e-01 4.5673e-01
92 418580 9.5417e-06 1.6718e06 4.7341e-01 4.5091e-01
94 453940 7.8608e-06 1.6198e06 4.27,87e-01 4.0343e-~1
95 469100 1.5610e-05 3.4350e06 8.9814e-01 8.6712e-01

~ ~ifi 5.2268e-03 4.7851e-03

(rb& that eo3= 10-3)

Table~O.3. Comparison of Exact and Approximate Failure

Probabilities. ~
 Sea . for pi for pi far pi for pi for i for
State ~j=0.5 Rj=o.7 ;;=0.9
j==o. o f~=o.1 ~j 0.3
(i)

00 4.31e-08 4.39e-08 4.58e-08 4.83e-08 5.14e-08 5.52e-H8

01 6.89e-08 7.05e-08 7.42e-08 7.88e-08 8.48e-08 9.24e-08
02 2.41e-07 2.49e-07 2.67e-07 2.89e-07 3.19e-07 3.57e-ti7
03 8.46e-07 8.80e-07 9.61e-07 1.06e-06 1.20e-06 1.37e-Ci6
04 4.07e-06 4.27e-06 4.75e-06 5.38e-06 6.19e-06 7.25e-06
05 2.15e-05 2.27e-L!5 2.56e-05 2.93e-05 3.41e-05 4.02e-d5
06 1.34e-04 1.42e-04 1.61e-04 1.84e-04 2.14e-04 2.51e-04
07 5.55e-04 5.88e-04 6.64e-04 7.57e-04 8.70e-04 1.00e-03
08 1.74e-03 1.84e-03 2.06e-03 2.33e-03 2.64e-03 2.99e-fJ3
09 4.13e-03 4.34e-03 4.82e-03 5.37e-03 5.99e-03 6.68e-03
10 6.25e-03 6.60e-03 7.37e-03 8.25e-03 9.23&03 1.03e-02
1.36e-02 1.42e-02 1.57e-02 1.72e-02 1.89e-02 2.06e-132
:; 3.21e-02 3.33e-02 3.58e-02 3.84e-02 4.lle-02 4.39e-02
13 4.92e-02 5.08e-02 5.40e-02 5.74e-02 6.08e-02 6.43e-L!2
14 5.98e-02 6.17e-02 6.55e-02 6.94e-02 7.34e-02 7.74e-02
15 9.00e-02 9.23e-02 9.68e-02 1.01+-01 1.06e-01 l.10e-01
16 1.30e-01 1.32e-01 1.37e-Dl 1.42e-01 1.47e-fll 1.52e-01
17 1.35e-01 1.38e-01 1.43e-01 1.49e-01 1.54e-01 1.59e-01
18 1.73e-01 1.76e-01 1.81e-01 1.87e-01 1.92e-01 1.97e-01
19 2.61e-01 2.63e-L!l 2.68e-91 2.72e-01 2.76e-01 2.80e-01
91 2.67e-01 2.70e-01 2.75e-01 2.80e-01 2.85e-01 2.89e-01
92 2.07e-01 2.lle-01 2.18e-01 2.24e-01 2.3t!e-01 2.36e-01
94 1.82e-01 1.87e-01 1.95e-01 2.03e-01 2.10e-01 2.17e-01
95 4.13e-01 4.14e-01 4.17e-01 4.20e-01 4.24e-01 4.29e-01

1.78e-03 1.83e-03 1.91e-03 2.02e-03 2.12e-03 2.22e-03

~ P~*fi
L
(NXe thateo3= 10-31

Table 10.4. Comparison of Failure Probabilities for Different

Correlation Co-efficients ( ~i: ) between Ms and Mw.
4
2E4
 2.3

2.2.

2.L
3
p *lo
f 2.0

1.9.

1.8.

1.7.

1.6.

1.5,
0. .1 3 .5 .7 .9

Jij

Figure 10.9. Effect of Correlation ( ~i; ) between Ms and Mw on

Lifetime Failure Probability (pf).

285
 $
,. 5
..
Linearized -- Exact Failure
Failure Sur~a.c/ . Surface.

Y
u

%N t
I

1 -1

%!! I -1

Figure 10.10. Exact and Linearized Failure Surface in Standard Space

(for Sea State 10 in Application Example 3).

286
 10.2.4. -- Application to Eighteen Ships Using Level 3. M.V.F.O.S.~
and the Improved First Order Methods:

The same eighteen vessels of the example application given in

Chapter 5 are used here to perform a comparison between Level 3,
M.V.F.0.S.M. and the improved method. The characteristics of the
ships are shown in Table 5.1 (Chapter 5 ). In Chapter 5 the
M.V.F.O.S.M. method has been applied to the eighteen vessels and
their safety indices have been determined on that basis. In this
example Level 3 and the improved first order (Hasofer/Lind and
transformation to normal variables) methods are also applied to the
eighteen ships. In Level 3, the following equation was used to
calculate the probability of failure (see equation 4.35):

Pf=[l-l[+]]+l[+-%l

+]+
e[-[ $x
@.6)

where v r and o r we the mean and standard deviation of the

. resistance, k is long-term mean value of the wave bending moment
(also equal to its standard deviation) and m. is the maximum
stillwater bending moment (considered deterministic). @ (,) is the -
standard normal distribution function.

In the improved first order method the Hasofer~ind Safety

index discussed in Chapter 5
was used (see equations (5. 15) to
(5.17). This procedure, however, yields identical results to the
M. V. F. O.S.M. method since the performance function is linear
(Mr - Ms -Mw = O). What results in a difference between the
M. V. F. O.S.M. and the improved method is the inclusion of the
distribution information as discussed in Chapter 5 (see equations
(5. 25) and (5.29)).

287
 The mean value of the wave bending moment k was
determined by a long term procedure described in the application
example of Chapter 5. Similarly, an initial yield limit state was used
to determine the resistance parameters as described in the same
application example.

Table lQ5 shows a summary of the comparisons of the safety

indices as calculated using the M.V.F.O.S.M. method and the improved
method. The safety indices were then converted to probabilities of
failure and the results are compared with the direct integration
method (Level 3) as shown in Table 10.6. Figure 10,11 shows the
absolute value of log I () pf for the eighteen ships and Figure 10.12
shows the probabilities of failure. III both cases the results are
platted versus ship length. Table10.7 shows the partial safety factors
Ar and At of the resistance and total bending moment, respectively,
as calculated from Chapter 6 .

Figures 10.11 and 10.12 show that, in general, the improved

method gives results closer to the direct integration method than the
M. V. F, O.S.M. method. This is solely because of fitting normal
distributions to the non-normal variables. However, the spread can
be quite large for some ships.

For these eighteen ships, both the improved method as well as

the M. V.F.O .S .M. method overestimate the safety of the vessel as
compared to the direct integration method, i.e., they err on the
nonconservative side.

The degree of approximation resulting from applying the

M. V. F. O.S.M. and the improved methods with respect to the direct
integration method varies considerably from one vessel to another
(see, for example, vessels no. 1 and 11 in Figures10 .11 or 10.12).
Inspection of Tables 1~5 and 10.6 reveals that the spread or errors in
these two methods as compared with the direct integration method
are strongly correlated to the total coefficient of variation of the load;
increasing with its increase.

288
 The lack of consistency in the degree of approximation when
using the M. V. F. O.S.M. or the improved method and the fact that they
lead to optimistic values of ship safety are matters of concern.

Vr Vc
ship t
P P
1 .13 O.aa 5.12 $.11
2 .13 .144 6.W 5.97

3 .13 .221 6.M 6.10

4 .13 .m7 4.70 4.64

s .13 .07s $.10 S.lo

6 .13 .074 5.2* S.27

7 .13 ,074 5.23 5.24

8 .13 .031 4.s7 4.a2

9 .13 .M S.20 5.1s

10 .13 .-7 4.M 4.02
11 .13 .140 5.2S 5.01
12 .11 .072 5.26 5.21
13 .13 .11s 5.51 5.43
14 .13 .110 4.W 4.37
1s .13 .072 4,48 4.48
M .1) .123 ],44 1.2s

il .13 . ?41 6.37 6.15

la .13 .133 S.60 5.49

 tnlp 4 P, (Wfow) P? (Iwmvd] Of (olr~t lnlqra~lom)

1 1 520[-7 1.611
(-7 1.541
C-7

z 6. OE-10 1.2(-9 $.Mf-*

3 1.5[.10 1.5[.10 4.390[-9
4 1.MIE4 1.501f.6 l.?b?r.b
5 1.6S9C-I 1.699(.7 1.172C-7
6 7.20M.7 6.825
[-8 1.2]][.7

7 a.4ac-a a.ol[-a 1.4652.7

8 $. Sac-? 7.178
[.7 1.784
[-6
9 9.965{-8 1.303
[-7 MU-l
10 5. M5E-7 7.178[:7 2. IW2-4
11 ?.60M-O 2.a52f-7 I.MM-b
12 }.21[-0 9.45[-8 1.wr-7
13 I .795E-a Z.sz[-a 1.1*7[-7
14 3. MC-6 6.212
[-4 1.%0[-$
15 3.732E-6 1,?32[-6 7.01
1[-6
16 2.ME4 5.770[-4 I.
OW[-J
17 I.U-lo 4.OC-10 7.430[-9
)a 1.075
[.0 2.01[-0 I.1mr-1

Table10.6

2 go

no
10

7
9

h 1
2

/
6
Is ,4
9t %
&

4
lb

3 -W
-,4,
I I I 1
Sog 6~ 7m am 1
9m Im Ilm
LEIP(flJ

l-d - 14.P.

Figure lQ 11

P,

10-.

IOti
-

u- -

10- -

m--

10- -

-*
10

m+ -

291

3!3
i)
 Table 10.7. Partial Safety Factors.

Bwd 011 bed on

MVFOSM Impr.wi Mcth~
Ship Ar d~ A. At
o.34a 0.370 1.13s
; ::Z&l ::2 0.267 1.391
3 1.352 1.596
4 0:407 :Z 1.156
0.349 i:% 0.366 1.116
i 0.326 6; 0.339 1.103
7 0.330 ;:: 0.343 1.103
!4 1.105 0.420 1.166
i :% 1.117 0.378 1.169
10 0.306 1.119 0.4Z3 1.190
11 0.349 1.219 0.430 1.354
0.39!6 L(M5 o.3i7 it
..099
1; 0.304 l.lti 0.350 1.252
1.153 0.510 L243
:: ;% 1.071 0.452 l.lti
16 0.596 1.1s2 o.6&2 i-263
17 i.ti 0.287 1.671
18 :%! 1.164 0.350 1.302

AP= strength
reduction
fatir.
At= load
~nification
factor.

10.2.5.
Implied in ABS Rules for Ship Longitudinal Strength

In developing a new code format, one should compare it with

existing practice to insure that the new method has some basis for
calibration, One way of doing this is to examine the reliability of
existing ships as was done in the previous examples for the eighteen
ships. This provides valuable information but has some limitations.
Among them are analyzing the ship in the as-built condition rather
than the code minimums, and the fact that in any large group of
ships, one is probably comparing different codes written at different
times.

Another way to tackle this problem is to simply investigate the

reliability implicit in the minimum strengths and loadings required
by the code. To accomplish this, the safety index ~ was calculated for
ten Series 60 ships, with Cb = 0.70 and L/B = 7.0. The minimum hull
strength and loadings required by the 1982 ABS Rules for Building

292
and Classing Steel Vessels were used to design the ships and the
following three assumptions were made:

1. The mean value of the hull strength was assumed to be

Ih = SM~ (107)

SM = section modulus from Section 6.3.l.a of the ABS

Rules
E = average yield stress of steel (assumed to be 31 ksi).

To determine the standard deviation, C~Vs of both 10 and 12

percent were used. As an example, using the 300 ft ship, the ABS
calculated section modulus is 6.243 X 103 in.2 -ft; then the hull
resistance is 8.640 X 104 ft-ton and the standard deviation is 8.640 X
I 03 ft-ton for a 10 percent COV.

2* The value of the stillwater bending moment was

calculated by Section 6.3.2.a of the ABS Rules. But this value was
considered to he an extreme value representing the 95 percent
exceedence level; that is, this value would exceeded only 5 percent of
the time. To find the corresponding mean and standard deviation,
this extreme value was used with COVS of 9.1 and 38.1 percent to
cover the range of possibilities suggested by ~0.9] and PQ10]. As an
example, consider the 300 ft ship for which ABS gives Msw = 2.662
X 104 ft-ton. Then the problem is to find Msw and 6SW such that

0.95 = FSW (2$62X 104) (lQ8)

and Oswjpsw = 0.091 or 0.381 where Fsw (.) is the cumulative

distribution function of the normal distribution. Using the standard
normal variate

293
it is easy to find from standard tables that t = 1.64 satisfies @8) and
using the condition for the COV provides a second equation. Taking
the 300 ft ship as an example, the 9.1 percent COV gives ~s w =
2.316 X 104 ft-ton and Usw = 2.108 X 103 ft-ton.

3. The value of the wave bending moment was calculated using

Section 6.3.2.b of the ABS Rules but again this was considered to be
an extreme value. The wave bending is known to fit the exponential
distribution where M = c, that is, COV = 100 percent. It remains now
to determine what exceedence level to assign the value derived from
the AB S Rules. This uniquely determines 1 for the exponential
distribution. From references ~0.9] and ~Ql 1], values of the expected
or average value of the wave bending moment h, based on rational
analysis, were obtained and then were compared with the wave
bending moment calculated from the 1982 ABS Rules. For example,
using the data from reference [lQ9] on the Mariner, k = 29 000 ft-ton
and Mw = 2.297 X 105 ft-ton from ABS Rules, one obtains from the
exponential distribution

P [ X ~ 2.297 X 105 ] = FX(X) = 1. e-(x/i)

= 1. e-(2.297 X 105/2.9 ~ 104 ) = 0.9996

so the exceedence level is about 0.1 percent. Similar results were

obtained from other examples and the exceedence level was set
somewhat arbitrarily at 1 percent, which is slightly conservative.
Following through on the 300 ft ship example with Mw = 3.569 X 104
ft-ton from the Rules, then

FX (3.569 % 104) = 0.99 = 1- e-3.569 X 104 /k

or L = 7750 ft-ton

Now the safety index ~ can be calculated using

P = (da-pH-?,
+ & + Ulu)/2 (lo.9)

294
{
p
 Equation (LQ9) is the equation for ~ using both the mean-value
first-order second-moment (MVFOSM) as well as the Hasofer/Lind
Method (without fitting a normal distribution), since the limit state
function is linear. Using the values of the parameters for the 300 ft
ship and noticing the pw = Cw = ~, then

P = 4.70 &QIO)

The partial safety factors can be directly computed using

equation (6.5). This results in the following values

As = strength reduction safety factor = 0.66

Asw = stillwater BM magnification safety factor = 1S08
Aw = wave BM magntilcation safety factor = 4.10 QQll)

It should be noted that these results depend on the validity of

the three assumptions discussed previously.
Although the values do
not change much with large changes in the coefficient of variation of
the stillwater bending moment, they are rather sensitive to the wave
bending moment exceedence probability (Assumption 2.) If an
exceedence probability, of 0.1 percent is used instead of 1 percent,
the following values result

~ = 5.44, AS = .0.56, Asw = 1.10, Aw = 3.94

instead of the values given by equations QO.1O) and (10.11). It is

therefore important to regard the values generated in this example
for the safety indices and partial safety factors as relative values of
the reliability implied in the ABS Rules rather than absolute values.
With this in mind, Table 10.8, columns 1 to 4 give the computed safety
indices according to the described procedure for strength COV of 10
and 12 percent, and for stillwater bending moment COV of 9.1 and
38.1 percent.

295 -
 Before calculating all the partial safety factors, one more
refinement was possible. This refinement is fitting a normal
distribution to the exponential wave load variate at the most likely
failure point. This is relevant since the calculation of the safety
index implicitly assumes that each variable involved has a normal
distribution and thus the safety index can be easily related to a
probability of failure.

The equation the most likely failure point of the wave

for
bending moment in the original space is determined from equation
(5.18) as

(1012)

The fitted normal distribution parameters can be now

determined using equation (1Q12) and equations (530) and (531 ) as
discussed previously. This leads to modified values -of the safety
indices according to the advanced Level 2 procedure. The results are
given in Table 10.8, columns 5 and 6, for stillwater moment COVs of
9.1 and 38.1 percent, respectively. Figure 10.13 shows plots of all the
results given in Table 10.8.

Finally, the partial safety factors using the equivalent normal

wave bending moment distribution were calculated. The
computation is straightforward according to the procedure described
earlier except for the partial safety factor associated with the mean
of the wave bending moment as obtained from the normal
distribution. This partial safety factor must be used in conjunction
with the fictitious normal distribution mean. Since the normal
distribution arises only as a part of the distribution adjustment
process, it is more relevant to determine the true partial safety
factor associated with the actual mean of the wave bending moment
(that is, the mean of the exponential distribution). This is done by
stipulating that the true partial safety factor, when multiplied by the
mean value of the wave bending moment, gives the same margin in
the checking equation as that of the normal mean multiplied by its

296
?08
partial safety factor; that is, both lead to the same safety index
values.

Table 10.9 provides a sumrhary of these results and the target

safety indices from which the partial safety factors where computed.
It should be emphasized that these results are meant to examine
trends and relative magnitudes rather than to be used in the
absolute sense.

In general, this analysis of the implicit safety in the ABS Rules

is somewhat surprising in two ways. First, the safety index ~, is very
consistent within each method over the range of ships lengths.
Second, the ~ factor decreases slightly with length while previous
results show it increasing. One possible explanation is that the
method of calculating the wave moment was not the same. Also,
since the previous analyses were done on as built ships, they would
reflect more factors changing than just the length. These would
include varying degrees of safety margin added by the designer to
the code-required minimum as well as different codes from different
years and classification societies.

297
 Table 10,8. Safety Indices of the Series 60 Ships

lap (j) (2) (3) (;) (5) (6)

Ship P P PP
300 4.705 4.726 4.231 4.335 3..0 3.648
; 4ca 4.650 4.@7 4.191 4.303 3.593
3 4.595 4.W4 4.160 4.269 3.311 3s30
4 z 4.M2 4,110 4.234 3.26s 3.487
5 4.489 :E 4.M6 4.196 3.227 3.438
6 z 4.4s 4.506 4.025 4.154 3.198 3.415
m 4.414 4.460 4.134 3.179 3.403
: 4.3MI 4.456 :&! 4.114 3.170
:E 4.367 4.433 3S3 ::L!& 3.160 :%
1: 12WI 4.346 4.411 3.944 3.149 3.373

(1)COV rmiatince lH COV mtillwabxbending9.1%

(2) COv raiatmce 10% COV mtillwamrbending 36.1%
(3) COV rmiatice C& COVatillwater bending 9.1%
(4) COV rmia-~ 12% COV stiilwshm bxling 36.1%
(5) COV rAaM.nce lfi COV stiliwabr kding 9.1% djusted wave &nding
(6) COV resistance 10% COV atillwstar tmding 38.1%adjusted wave binding

O STILLWATER C.O.V. = 38.1 %

P A STILLWATER C.O.V. *9.1%
5.0 MVFOSM METHOD
STRENGTH C.O-V. ~ IO %
$

45

40
ADvANCED METHOD
MVFOSM METHOO
STRENGTH C OV. IO%
STRENGTH C.O.V * 12 %
b -

3.5

3.0
t
I I 1 I I 1 1 I I
2.51
300 400 500 600 700 800 900 I000 1100 I200
L. B.F! (FT)

Figure 10.13. Safety Index (ABS) versus length between

perpendiculars of Ships

298

w
TabIe10 .9. Partial Safety Factors and Safety Index of Series 60 Ships

w 0.s605 1.031 6.515 3.420

; 400 0.6662 1.029 6.434 3.364
3 0.8710 1.026 6.34a 3.311
4 : 0.8750 1.027 6.261 3.265
5 700 0.E7M 1.027 6.165 3.227
6 600 0.6763 1.027 6.117 3.196
7 WO 0.M07 1.027 6.06$ 3.179
0.6606 1.027 6.046 3.170
: i% 1.027 6.018 3.160
10 12(KI :E l-on 5.9e$ 3.149

All factom hod on mength COV - 10%ud stillwater moment

Cov - 9.1%.
(1) Equiv4fent partisl mefetyfmtm, x
(2)Tsrget safety indicm from which dFtidwaemOmenL
putmlmfety factors
e.re
derivd

299
10.1 Lartkford, W., The Structural Design of the ASR Catamaran
Cross Structure, Naval Engineers Journal, Aug. 1967.

10.2 Hogben, N. and Lumb, F., Ocean Wave Statistics, National

Physical Laboratory, Ministry of Technology, London, 1967.

10.3 Pierson, W. and Moskowitz, L., A Proposed Spectral Form for

Fully Developed Wind Seas Based on the Similarity Theory of
S.A. Kitaigorodskii, Journal of Geophysical Research, Vol. 69,
Dec. 1964.

10.4 Loukakis, T., Computer Aided Prediction of Seakeeping

Performance in Ship Design, Report No. 70-3, Department of
Naval Architecture and Marine Engineering, MIT, Aug. 1970.

10.5 Chen, Y, K., Kutt, L. M., Piaszczyk, C.M., and Bieniek, M-P.,
Ultimate Strength of Ship Structures, Trans. SNAME, Vol. 91,
1983.

10.6 Mansour, A., Methods of Computing the Probability of Failure

under Extreme Values of Bending Moment, Journal of Ship
Research, Vol. 16, No. 2, June 1972.

10.7 Beck, R., A Computerized Procedure for Prediction of

Seakeeping Performance, Report No. 69-2, Department of
Naval Architecture and Marine Engineering, MIT, March 1969.

M8 Ferro, G. and Mansour, A., Probabilistic Analysis of the

Combined Slamming and Wave-Induced Responses, Journal of
Ship Research, Vol. 29, No. 3, Sept. 1985, pp. 170-188.

10.9 Mansour, A. E., Probabilistic Design Concepts in Ship Structural

Safety and Reliability, Trans. SNAME, Vol. 80, 1972.

300
10.10 Soares, C.G. and Moan, T., Statistical Analysis of Stillwater
Bending Moments and Shear Forces in Tankers, Ore and Bulk
Carriers, Proceedings, Third Iberoamerican Congress of Naval
Engineering, Madrid, 1982.

10.11 ?Iorne, M.R. and Price, P. H., Commentary on the Level-II

Procedure, Supplement to Rationalization of Safety and
Serviceability Factors in Structural Codes, Report 63,
Construction Industry Research and Information Association
(CIRIA), London, Ott. 1976.

301
11. CONCUOXNG RIMRKS AND RKCM!ENDATIONS

The powerful tools of the theory of probability provide

excellent means for assessing the safety of marine structures under
certain conditions of uncertainty. Probabilistic methods have been
developed and used in practice for describing the random loads
acting on a marine structure and the uncertainties associated with its
true strength. The safety margin between an extreme loading
combination and the strength of a structure is then assessed through
reliability indices and probabilities of failure. Such reliability
analysis is only a small but important part of the total probabilistic
approach for designing or checking a marine structure. Partial safety
factors and safety formats suitable for use in design and for
implementation in Codes and Rules have been advanced and used by
practitioners as well as Classification Societies and Code Developers.
Fatigue analysis has been developed in several reliability formats
which allow for the estimation of the probability of failure from the
load history and the fatigue strength of the material. Complex
redundant structures and structures with multiple failure modes or
mechanisms have been treated using system reliability concepts
which are being rapidly developed at the present time. In short,
powerful and sophisticated probabilistic tools are currently available
for use by the marine industry. There are, however, several
shortcomings that plevent a wider use of the probabilistic and
. .
reliability methods in the design process. These include:

1. Use of reliability in checking and

analysis design processes
requires more information on the environment, loads and the
properties and characteristics of the structure than a typical
deterministic analysis. Often such information is not available or
may require considerable time and effort to collect -~. Time and
schedule restrictions on the design are usually limiting factors on the
use of such sophisticated methods.

2. Application of probabilistic and reliability methods usually

require some familiarity of basic concepts in probability, reliability

.
302
and statistics. Practitioners and designers are gaining such
familiarity through seminars, symposia and special courses.
Educational institutions are also requiring more probability and
statistics courses to be taken by students at the graduate and
undergraduate levels. This, however, is a slow process that will take
at least one generation in order to produce the necessary infra-
structure for a routine use of reliability and probabilistic methods in
design.

3. The two shortcomings stated above are not severe drawbacks

in connection with development of Codes and Rules based on
reliability analysis since such a development requires a one time or
a more consolidated effort to collect the necessary information. In
addition, the one time code format and development can be done
by experts in the field. But here the inertia of tradition comes into
play which makes any new approach, reliability or otherwise,
difficult to incorporate. This, however, has been changing and more
Classification Societies and Code organizations have taken an active
interest in the probabilistic methods and developed Rules and Codes
based, at least partially, on reliability.

4. On a more technical aspect, the reliability analysis did not

deliver what it initially promised, that is, a true measure of the
reliability of a structure by a true and actual probability of failure.
Instead what it delivered is notional probabilities of failure and
safety indices which are good only as comparative measures. Only
notional values are delivered because of the many assumptions and
approximations made in the analysis producing such probabilities
and indices. These approximations, deficiencies and assumptions,
however are made , not only in the reliability aspects, but also in
other aspects and disciplines used in the design. Such aspects
include determination of loads using hydrodynamics theOry and
approximations made in the structural analysis and response to the
applied loads. When all such assumptions and deficiencies are
removed from the design analysis, the resulting probabilities of
failure will approach the true probabilities.

303
 In spite of the shortcomings stated above, use of reliability
analysis in design provides advantages and unique features. These
include:

1. Explicit consideration and evaluation of uncertainties associated

with the design variables.

2* Inclusion of all available relevant information in the design

process.

3. Provides a framework of sensitivity measures.

4. Provides means for decomposition of global safety of a

strue ture into partial safety factors associated with the
individual design variables.

5. Provides means for achieving uniformity of safety within a

given class of structures (or specified nonuniformity).

6. Minimum ambiguity when updating design criteria.

7* Provides means to weigh variables in terms of their

significance.

8. Rational guidance for data gathering.

9. Guidance in novel designs.

The advantages seem to outweigh the drawbacks and it is

almost inevitable that the probabilistic and reliability aspects will be
used in designs where randomness of the variables is an important
consideration. Based on these conclusions, the following
recommendations am made:

1. The major effort currently progressing in the development and

application of reliability methods to marine structure should be
continued and expanded. Such efforts will not be wasted since,
most likely, some of the developed procedures will, sooner or
later, be used in design.

304
2. In the calculations of the reliability indices and probabilities of
failure, the resulting values depend considerably on the
methods used in determining the loads acting on the strut ture
(e.g. extreme versus long-term waves loads) and on the method
of combining these loads. A need exists for standardizing
such procedures for use in design.

3. A study of target reliability based on existing ships or

minimum Rule requirements for the primary strength should
be undertaken based on such a standardized load procedure.

4. Studies and additional development of reliability methods are

needed for the secondary (stiffened panels) and tertiary (plates
between stiffeners) aspects of ship design.

5. There is currently a strong tendency to neglect level 3

reliability analysis in favor of level 2 because of the difficulties
stated in Chapters 4 and 5 of the report. Certain
simplifications can be made however within level 3 framework
which would make it possible for application to marine
structures. Such simplifications and further developments of
level 3 are worth persuing. Similarly, application of simulation
techniques should be further studied.

6. System reliability is an essential aspects of reliability analysis

of highly redundant structures such as offshore platforms.
Additional work is needed in this area particularly in regard to
simplifying and reducing the number of permutations of
possible failure paths and the corresponding c~mputations.

305
 APPENDIX 1
HELPFUL INFORMATION

In this appendix some useful infomuation on several aspects of reliability of

marine stmmtures are described. They relate to topics which appeared in various
chapters in this report where reference is made to this Appendix.

. ..
Weibull Tlstributi
AiA~Aq 1

The probability density fmction (p.d.f.) and cumulative distribution

fiction (c.d.f.) of the Weibull Distribution are given by

x q-l
p.d.f. = &(x) = (V@
()
~ e-(X/k)~ (1)

c.d.f. = Fx(x) = 1 . ~-(x/k)< (2)

where k and ~ are parameters to be determined from data, e.g., data of wave
amplitude or wave-bending moment amplitude. .

The first two moments (mean and variance) of the Weibull distribution are
given by:

Mean = E(x) = kr(p+l) (3)

Variance = Var(x) = kz ( r (2..i-i+ 1) - (r&l + 1))2) (4)

where Ht) is the Gamma function defined as

The Gamma function is tabulated in many Handbooks, e.g., Handbook of

Chemistry and Physics.

Some properties of the Gamma function are described as follows:

306
 r(t) = w rm
r (t+l) = t r (t) = t (t-1) . . . (t-r) 17(br)
i.e., generalization of the factorial function.
r(l) = 1
r (n+l) = n! for any n = integer
1
= Tx
z ()
rn+~
() .
1X3X5 . . . (2n-1)
2n T z; n = integer

rn+~
() s
lx3X5...
211/2
(1)l)
r
~ .
$
n=evenhtig~r

The Weibull distribution reduces to two important special cases as follows:

9. Exponential Distribution

When != 1 and k = 1, the Weibull distribution reduces to the exponential

distribution with parameter L From equations (1) and (2), the resulting p.d.f. and
c.d.f. are

(5)

Fx(x) = 1 - e-fi X>o (6)

From equations (3) and (4), the mean and variance of the exponential
distribution are given by

E&) = Ir(l+l) R ~r(2) = L (7)

Var &) = ~2 (r(z+l) . r(2)2) = ~2 (24) = ~2

(8)

b
~~~

Whenl=2andk. ~E the Weibull distribution reduces to the Rayleigh

distribution. Notice that, if the Rayleigh distribution is resulting from a
stationary Gaussian process as the distribution of the peaks, then E as defined

307
f
37!
here is the mean square value of the process, i.e., the area under the spectral
density of the process. From equations (1) and (2), the Rayleigh distribution is
given by

p.d.f.: $x(x) = e-~2121? X>o (9)

:

c.d.f.: Fx(x) = 1 - e-X2fi X>o (lo)

and, horn equations (3) and (4), its mean and variance are given by

E(_Xti=~r#+l=~&&= ~&E (11)

()

VaI(XR) = 2E(r(2) - (r(++l)~)= ~ (1-$) o-z)

Ei
~~ nfhi ri

Several methods can be used to estimate the Weibull distribution

pa~ameters from a set of data. Since the exponential and the Raylei~h
distributions are special cases, similar methods can be used to estimate their
parameters. The methods include the method of moments, Weibull probability
paper, the maximum likelihood method and a method based on order statistics.
Only the moment method and the Weibull probability paper are discussed here.
The advantage of the probability paper over the moment method is that it provides
a mean for checking if the Weibull distribution actually fits the data or not as will
be discussed later.

a Method of Mometi

The mean and standard deviation of a data sample can be determined from
the usual equations:

308
 i~l
ax = Sample slandard deviation =
( L
~
=
(q-Rp
)
*

The resulting values for ~ and ax can be used in conjunction with equations
(3) and (4) to determine the values of k and !, or more conveniently from the ratio:

(13)

which is a function off only. Thus ~ can be estimated from (13) and then inserted
in equation (3) or (4) to determine k.

b. Weibull Probab ilitv PaneK

The Weibull distribution function is given by:

Fx (X) = 1. ~ -(xik):

therefore,

log log ( 1- Fx (X)) = -! log (A)

or,

(1-M=~ogx ogk
1oglOg
Insert

w = log log
(~) ; v ogx

the linear relation results

(14)

309
 So if (l-FX(X))-l or (l-FX(x)) is plotted against x on log log versus log paper a
straight line is obtained (if the data fits the Weibull distribution closely). The slope

of the straight line is ~ or -1, respectively, and the intercept with the axis is -1log K.
Thus k and ~ can be determined.

Notice that, in addition to providing a mean for estimating the parameters

k and 1, the WeibuIl paper is usefid in examining visually the quality of the fit.
Goodness-of-fit tests such as Chi-square, W-statistics and Kolmogorov-Smirnov
tests may also be used to examine more accurately the quality of the fit and to
determine which of several candidate probability distributions fits the data best.

kB 4.=1
1= 1,W

Figure AL Weibull Plot of SL -7 Five Year Data

310

jw-
More detailed information on these tests and examples of their use in conjunction
with wave bending moment data fkom SL-7 containerships are given in reference
[Al]. A sample of Weibull probability paper which shows the fit of 12319 data
points from SL-7 containerships is shown in Figure Al obtained from reference
[Al].

A1.2. The Safetv of Index V ersus Probab ilitv of Failure for Normal and Othe~
Distributions;

In Chapter 4 and 5 it was shown that, for the simple margin M (or limit
state function) given by

M=g(xl, xz) = S-Z (15)

the probability of failure pf is given by (see equations 4.4 and 5.5)

.
w = ~FS(z), fz(z)dz (16)
o

= FG (-~) (17)

where FG (.) is the cumulative distribution function of the standardized Margin

G (see equation 5.4) and ~ is the safety index defined as the margin mean
divided by its standard deviation. If S and Z are both normal, then the margin M
defined by (15) is also normal and FG (.) becomes the standard normal cumulative
distribution function tabulated in many handbooks. Thus the relation between pf
and ~ can be easily computed.

Figure A.2 obtained from reference [A.2] shows the relation between pf and
~ for some other distributions of S and Z and spetied values of their coefficients of
variation ( VS= 0.13 and VZ = 0.10). The plot shows that pfis sensitive to the type of
distributions of S and Z in the higher values of ~ (range of low probabilities of
failure).

311

32j
 1
S - NORNAL //
v~ = 0.13 z - Exponential
+ NORMAL
v= = 0.10
10-7. s- NORMAL
z- ASYMPTOTIC
EXTREKE II

10-6:_
s.-NOWAL
2- EXTRH4E(n=3) + NO

10-5L
s- k3YMPTOTIC
EXTREME I
z- ASYMPTOTIC
EXTRHfE
10-4~

wvev~~

Nott4k L
10-3=

10-2

B 1 I I
L 2 3 4 5 6
SAFETY INDEX $

Figure A.2. Probability of Failure Versus the Safety Index

312

32$1
 REFERENCES

A. 1. Mansour, A. E., Jan, H. Y., Zigelman, C. I., Chen, Y. N. and Harding, S.

J., Implementation of Reliability Methods to Marine Stmctures, Trans.
SNAME, Vol. 92,1984, pp. 353-382.

A.2 hkmour, A. E., Approximate Probabilistic Method of Calculation Ship

Longitudinal Strength, Journal of Ship Research, Vol. 18, No. 3,
September 1974, pp. 203-213.

313
Appendix 2: Computer Program CALREL for Performing Reliability
Analysis

A Brief Description of CALREL - A Computer Pronam for

Component Reliability Analysis:

CALREL is a batch processed computer program in FORTRAN

language suitable for execution in both mainframe and
microcomputers. Given a probabilistic characterization of the basic
random variables, and an analytic performance function (limit state
equation), the program calculates the Hasofer-Lind reliability index,
13HL, in the standard space of uncorrelated variables (U space). The
program calculates the probability of component failure if
probability distribution (level III method) of basic physical variables
are provided. The output includes sensitivity measures of the
reliability index and probability of failure with respect to basic
variables, deterministic parameters in the performance function and
the distribution parameters. Following is a brief description of
special options and features of CALREL. The attached Users Guide to
CALREL is a self explanatory document of all the other options and
features. .-

Input Description: .

The input to CALREL consists of two parts: i) input data and ii) user
provided subroutines. The data input defines basic physical random
variables, i.e., their mean standard deviation, correlation, etc. and, or
parameters of the optional distribution functions. For level-II
methods only second moment characterization of random variables
are necessary. Both mean value first order second m o m e n t
(MVFOSM) and Hasofer-Lind first order second moment (FOSM)
reliability index can be calculated. For level 111 characterization of
the random variables two options are available: i) first order
marginal distribution method [FOMD, ref .A.1] or ii) first order full
distribution method [FOFD] using Rosenblatt transformation.

314
M
If the basic random variables are independent they can be specified
completely by their respective marginal distributions. For
dependent variables if only marginal distributions and pair wise
correlation among them is known then, a (non-unique) joint
distribution model is implicitly assumed that is consistent with the
specified marginal distributions and correlation structure [FOMD,
ref. All. The marginal distribution function can either be chosen from
the program library or can be specified in an analytical form through
a user specified subroutine called df, The correlation structure for
dependent variables is specified in the input data section.

For full distribution method [FOFD] using Rosenblatt transformation

the following conditional distribution functions are analytically
specified through a user defined subroutine called hfun. [See
Example 3 in Users Manual.]

Hi(xi I Xl, .... xi.1) = p(xi~ xi I Xl = xl, .... xi-l = xi.1) (Al)

If f~~ and F~(~, respectively, represent the joint density and joint
distribution function of & we have:

Xi

Hi(xi I Xl, .... xi-l) =

JXilll ,.....l(xilxl~xl~ c, xi-l) dxi

ai-1
Fxl, ....Xi(xl ....xi)
= axl .....~Xi-l
(A.2)
Xl, .....li.l (Xl,..., Xi-1)

The program then implicitly uses the following transformations

between the basic variables space and the standard normal
uncorrelated space :

315
 U1 = @-l[Hl(Xl)] = @-l~xl (Xl)]

U3 = @-1[H3(X3&X2)] (A.3)

Un = @-l[Hn(XnlXl, .... Xn.1)]

In addition to the above specifications in the subroutine hfun the

standard deviation of the basic variables are required as input data
for calculating iteration steps in the optimization scheme to find the
~ -point. The user can also specify parameters of the above
conditional distributions (h-functions) through input data.

Performance function: The performance function (limit state

equation) in CALREL is specified analytically through a function
subprogram g. The subprogram returns a value of the performance
function for each call from the main program specifying a value of
the basic variable & The parameters for the analytic performance
function can be passed from the main program, if defined, through
the input data.

The main program uses a finite difference scheme to calculate the

gradient vector of the limit state surface at the iteration point.
Hence, if an analytic performance function is not available the
subprogram g can be made to call other programs (e.g., finite
element, dynamic analysis program, ctc.) to return a value of the
performance function. Since finite difference scheme is used to
calculate the gradient vector at the ~ - point, a number of
performance function values may be required involving great
computational efforts. It is desirable to be able to input the gradient
vector directly when
available (either analytically or through
numerical values returned by other programs such as finite element
etc.). This is beyond the capability of the present version of CALREL
but can easily be achieved through minor modification of a
subroutine in the main program. The user provided subroutines are
provided in a file called user.for which is compiled and linked to the
main body of CALREL, each time a new problem is solved.

Output Description

The output of CALREL consists of reliability indices, probability of

failure for level-111 analyses and various other sensitivity measures.
In level-III analyses probability of failure results can be obtained
based on both first order (tangential hyperplane) and second order
(quadratic hypersurface) approximation of the limit state surface.
Two different second order approximations to the actual limit state
surface are available, based on point fitting and cuwature matching
procedures. For an approximated quadratic hypersurface the
probability content is calculated by four different approximating
formulas. The different sensitivity measures calculated by the
program can be described as follows:

(A.4)

where Q* is the design point (or ~-point).

& is a sensitivity measure of ~ HL with respect to the standard

variates (UI, U2, .... etc.).

2. Measures of sensitivity with respect to basic variables ~ at a point

x * (corresponding to Q* in Q space) is given by

vp(&*)
=~r D-1 (A.5)

317
 where ~ = diagonal matrix of standard deviation

and, 1 = L-1, R = ~ ~T and ~ = correlation matrix.

To make variations ~xi*; i = 1, .... n; equally likely V ~(x*) is scaled by

the corresponding standard deviations, i.e.,

A unit sensitivity vector is now defined as:

(A.6)

Gamma (y) is a relative measure of importance among basic random

variables.

The program also calculates delta and eta normalized (each

variation equally likely as in ~ and y) sensitivity vectors with respect
to the mean and standard deviation of the basic physical variables.
If desired the program also calculates sensitivity measures of
reliability index and probability of failure with respect to other
distribution parameters and deterministic parameters of the analytic
performance function.

References

A. 1.Der Kiureghian, A. and Liu, Pei-Ling (1986). Sbuctural Reliability

Under Incomplete Probability Information. Journal of Engineering
Mechanics, ASCE, Vol. -112, No. -1.

318
 ,

.,

.,

. .
User% Guide to
C6LREL - .
-h+
.
A First and Secfind-ctrder Structural Reliability &nalysis Pragram
:.
-. G
.-
INPUT DATA

1. Title --- Format fA80)

CCI 1umn Variable Descripticm
1-00 TITLE Alphanumeric
- - deseriptian CIT zne praDlam

2. Cclntral Data --- Format (515/71!5/315,4d10.0)

1- 5 1FQ Type c,f technique used
IFO=l tlVFOSM d
IFO=2 FOSM
IFO=3 FOMD
IFO=4 FOFD {Rclsenblatt transfarmatien)
6-ICI Iso Type af =ecc~nd-erder apprclximation
1S0=0 Fir~t-6rder analysis anly
1S0=1 Point fitting m~thod
1S0=2 Curvature fitting methed
1S0=3 Bc{th Paint and Curvature fitting=
11-15 ITG Type of integration schemes used in sec~nd
order appr~ximaticm
ITG.eq.O Breitung formula, Tvedt 3-term
farmula and Tvedt single integral
ITG.ne.O All abave three schemes plus
Tvedt dcmble integral
16-20 ISv Type Of sensitivity analysis required
I SV=Q Na sensitivity analysis
ISV= 1 distribution parameters
ISV=2 performance functicln parameters
ISV=3 distribution and performance
-functian parameters
IRS restart code
IRS.ne.O restart analyzing n old,
uncanverged prciblem
IES.eq.6 analyze a new prcsblem
1- 5 NX Number ctf basic variable=
6-10 w Number @f deterministic parameters in
the performance function
NU Number of user-pravided distributions
NS Number @f parameters in user-defined full
di6trib~ti~~S (Gppllcable when IFO=4)
NCORR Flag for cfirrelati~n matrix &
(fipplicable when IFO.ne.4) ;
tJCCRR.@qzC) Unesrrelated vqriable~
NCDRR.ne.O Correlated variables
26-30 INIT Flag far initialization
INIT.=q.O Start frem mean paint
INIT.ne.O Start point specified by user
31-3!5 I FR Output cede
IPR.eq.O Output all iteratian steps
It%?.ne.O Output at every ipr %tep~
?lq ?1(
 1- 5 10PT Type of optimizati~m scheme used
IOPT=l HL-RF methbd
XOPT=2 Modified HL-RF method
IOFT=3 Gradiemt %ojectien method
6-10 NIT1 Maximum number of iteratian cycles
Default=100~ Maximum=100 -.
NIT2 Maximum steps in line search:
Default=4
TOL Convergence talerance
Default=O.OOIY Minimum=O.001
OPT1 Step size reductien factor in line s=arch
IOPT=lZ t)efault=l.O
IOPT=2 ar 3- Default=O.5
0FT2 Optimization parameter
IOP1=2 Parameter c in descent functic~n
Default=10
IOPT=3 Ccinvergence tolerance fc~r line
search; Default=TOL
0PT3 Optimization parameter
IOpT=3 Maximum step Size in line search
Default=4.O

3. Usm-d~fined Distribution --- Format (13!A20)

~~ip this sectinn if NU=O.
Fclr each u~~r-defined distributl~n input:
1- 5 N13 I SU Type number c!f ue.er-defined distribution
NDISU > 20
6-25 UNMIE Name of user-defined distributicin.

Hasic Randc#m Variables

Skip this- sectign if IFO=4. .
Fctr each basic .variable with NDIS < 21: (213,5D1CI.0)
1- 5 W Variable number
5-10 ND h Nnxs = abs(ND) : Distributi~n type.
NDIS=i Normal
NDIS=2 , Lclgnarmal
ND1S=3 Gamma
NDIS=4 Shifted Expc~nential
NDIS=!5 Shifted Rayleigh
ND1S=6 Uniform
NDIS=7 B=ta
NDIS=ll Type-I Largest Value
N91S=12 Type-I Smallest Value
- NDIS=13 Type-II Largest Value
NDIS=14 Weibull
11-~cl Pi Distributicm parameter 1
ND>O P1 : mean value -
ND<O P1 : a= defined i; Table 1
21-30 P2 I)i%tributien parameter 2 -
ND>.0 Pa : Standard dev~:ation
ND<O P= z as defined in Table 1
31.-4<1 P3 Di!atributi~n parameter 3
4 i-!50 P4 Distribution parameter 4
Presently~ P3 and P4 are applicable
only when NDIS=7.
51-66 XINIT Initial value of x; Only needed when INIT=O


Far each basis variable with NDIS > 20: (215~5DlC).O/15,2D10.0)

i- 5 NV Variable No.
6-10 NDIS Distribution type
11-20 Pi Distributicln param~ter 1 :
21-30 F2 Distribution param@ter 2
e 31-40 P3 Distributicln parameter 3 ~
Ll
>..
,= 41-50 P4 Distribution parameter 4 *-
.-. 51-60 XINIT Initial value of x; must be defined
even if INIT=O
1- 5 IB Flag ffir beunds
IB=O NCI bounds
lB= 1 Has lewer baund
IB=2 Has upper bcund
IB=3 Has lc,wer and up~er b~und%
6-15 END 1 Lawer bc,und of the basic variable
Applicable when IE=1,3
14-2!5 END= Upper bound uf the basic variable
Applicable when IB=2,3

~. Basic Rar,dam Variable- --- Farmat (15,2D10.Q, 13,2D10.C~)

Ship this secticin if IFO=l, 2* or 3.
Fcir ~ach basic variable input:
1- 5 NV Variable NQ.
.
6-15 SIG Standard deviatian
1&-25 XINIT Initial value clf x
24-30 IB Flag far bounds
IE=o Na bcwnds
IE=l Has lawer bcwnd
IB=2 Has upper bound
IE=3 Ha= lciwer and upper bounds
31-40 13ND1 Lower beund af the basic variable
Applicable when IB=2,3
41 -!5CI BND2 Upper bound of the basic variable
Applicable when IB=2,3

6. Correlaticln Matrix --- Format (*D6.0)

Skip this section if NCORR=O.
1-1o RO L6wer triangle Of the correlation matrix
1-10,11-20 excluding the diagonals. Read it raw-wise
.. . and in triangular shape.

7. Parameters in Full Distributim Functicm --- Fclrmat (8D1O.C})

Skip this sectian if NS=O.
l-lCI~... DS Values Of the parameters in the full
-, cumulative distribution funtti~n
--
..
~8. Parameters in Performance Functi@n --- Farmat (6I31O.6I
g Skip this section if NP=O. p
l-lC}~... DP Values @f the deterministic parameters in
the performance functian *:

321
 9. Contrcll Flag fer Prclgram Executictn --- Fc4i-mat (15)
1- 5 NEXT Cantrc,l flag fclr program executian
NEXT=O stctp execution .
NEXT=l Restart a brand new analysis
Gll th~ afclrementiclned data
E shc~uld be input After this line.
~Lg
y NEXT=2 Re-analyze the oIU problem
.*. with a different set of parameters
in the performance function.
Only the values clf the parameters
in the perfcirmance function shcluld
be input after this line.

(1) If the nearest puin~ is not fctund in NIT1 steps- the final
status of the analysis will be stored in an unfc~rmatted
rile calrel.sav. This file mu~t remain unaltered if the
analysis is to be continued in an ensuing run.

(2) In a restart prablem? the pragran reads cmly the title and
the contrcll data frclm the input. file. The initial status
c,f the prctbl~m are read from ealrel.Sav*.
In ctrder te be ccmsistent, IFO, NX, NP, NU) N~? and NCORR
must be the same as the previ~us run.

(3) la averride the r==trictian~ NITI<1OO and TOLt.C~.001y input

negative NIT1 and TOL. Their ab=cllute value= will be used
in the analysis r~gardless af the limits.
. i
.

.. - . -..

PF

p\s A 70
x*e
, ..
t

X=*

I ,

P2 = da >0

t3 X78
,..pz
=t=70

pl=u, -
-::
X7V ,
pz=v72

t:

I
. .
.

323
 functlcin g(xsdp)
c
c..~..functicm subroutine te c~mpute the limit-state functian
C...*.X = vectar of basic variabl~s
C.i...dp = vectc~r of deterministic parameter%
c
implicit r~al@ (a-hla-~)
dimensictn x(l)~dp(l) .
. .

.
:.
9 E- ;
re**rn +
en~
.
=ubtioutine df(par?x~nd,cdf?pdf)
c
c . . . ..Subrcmt~ne tn cclrnpute pdf and cdf of user-defined distributions
c.....par = vectar ctf parameter distributions
C.....x = value of variable
c.....nd = distribution number (>20)
c.....cdf = computed cdf value
c.....pdf = camputed pdf value
c
implicit realw8 (a-h$a-z)
dimensien par(41
g= to (10?20,.. .) nd-20
10 calf= ...
pdf= ...
return
20 cdf = .. .
pdf= ...
return
.
.
return
end

subrclutine ,hfun(x~ih?ds~hi)
c
c.....Subreutine to compute conditional CDFS far Rasenblatt transformatic!n
c.....>: = v~ctcjr clf basic variables
c.....ih = rc,w number in Ra5Enblatt ~ransfarmaticsn
c.-...ds = v~ctclr cif deterministic distribution parameters
c.....hi = value of ith conditional cdf
c
implicit real+8(a-hsa-z)
dimension x(l),ds(l)
go to (10~20~...) ih
10 hi = ...
return
20 hi% ...
re*rn
*
~. =
.
return
end

324
 k41e Usef*+Qr
.
furtctian g(x!dp)
implicit real*8 (a-h to-~)

dimensicm x(l), alp(l)
9 = dp(2)*x (2)** 2-alp(l)*>:(l)
return
end

subrc~utine dftpar?):?nd?cdf!pdf) ,.
implicit real*S (a-hSo-z)
di~n~icm par(4)
-. re.mrn
: end:
.

=ubroutin~ hfun{x?ih?ds~h)
implicit real+=(a-hsa-z)
dimen~ion x(l),ds(l)
return
end

r-

example 1
331 3 c1
2 2 00001
1 20 8 -C).c)ool
1 -11 90.9992 .CM41275 0. 0.
2 2 at>. !5. (). c1
1.0 0.!5
c1

0.

325

. ---- ---. --.

.-.---
 . ***++*+******************4****************+*****************
* Un i ver s i t y ctf Cd 1 i for ni a *
* Department of Civil Engineering *
* Divisian of
* Structural Engineering and Structural Mechanics *
* *
F OS R AP

g First (kder Structural Reliability Analysis Prctgram +

$ & *
*. I)e\elopeclBy *
* Pei-Ling Liu and Rrmen Der Kiureqhian *
* Last Revision: September 1986 *
*
* Extended far F~FD and SOSWP by *
* HOtJG-Z~NS LIN, March 1~=~ *
+****+***++*+**************+********************************

+*** input data -- prebl~m 1 ****

exaample 1
type clf first-clrder technique used .. ... ..ifa= 3
ifca=i ...mean-value! 1st order, 2nd moment (mvfclsm) method
ifc,=~ ...first-c,rder? seccind mc~ment (fctsm) methc~d
ifc,=3 . ..first-arder. marginal distribution (fclmd) methad
ifct=4 ...first-c.rderl full dist.(flasenblatt trans) method
type af secclnd-nrder technique used ... ..is~= 3
isct=~ ........ ... ..nci secc~nd-arder approximation
lsu=l . . . . . . . . . . . . . . . . . . . . . . peint fitting method
1sc~=2 . . . . . . . . . . . . . . . . ..curvature fitting method
i ~nz~ . . . . . ..point and curvature fitting methods
type c~f integration s.cfiemesused in seccmd-arder
analysls . ......... ... ..... ..... . ..... . . ltg= 1
itg.eq.C~ ....Hreitung? Tvedts 2-term and single integral
itg.ne.tj -.-all c.f th= abclve plus Tvedtss dauble integral
ty~e of sensitivity analysis required ...isv= 3
isv=Cl . ............... ..... net ~ensitivity analysis required
lsv=l .. ....... ... ..senzit ivity clf distribution parameters
i5v=2 . .. ....sensitivity cff perfclmance function parameters
isv=s . . ..dlStrlhltiCttT and perfclrmance functian parameters
number ctf randctm variables .... .. ..... ....nx= 2
number Clf dc~terministicparameters ... . ...np= 2
num5er clf us~r provided distributicm . ....nu= 0
mumtler Of pdramet=r= 1~ u=~r-defined full
di~tributians ....... .. .......... ..... ....ns= c}
ccirrelaticcn structure . . . . . . . . . . . ... . . . .ncc~rr= (1
nCctrr.eq.O . . . . . . . . . . . . . ..uncc!rrelated variables
ncctrr.n_e.0 . . . . . . . . . . . . . . . . . carrelat~d variablras
. .
initlall~~~tlctn flag ............. .. ... ..init= 6
init.eq10 ... .. ... =.initialization at mean paint
init.me.~ ...... .... ......init ialization by user
.
Clutput flag . .* . . . . . . . . . . . . . . . . . . . . .m. . . ipr= 1
lpr.eq.O . . . . . . . . . . . . . .cutput anly final results
ipr.ne.~ . ....= . .... ...autput at every ipr ~teps
.-.

326
 Icop T=l . . . . . . . . . . . . . . . . . . . . . . . . . . . ..nr-flkli,~ *,iQu

,iclpt=2 ....................Mc.dified
RF-HL method
iopt=3 ...............Gradient Prejecticln method
maximum number ef iteratian cycles .....nitl= 20
ma%imum steps In lzne search .... ...... .ni t2= 4
ccfnverqence. tolerance . . . . . . . . . . . . ..tal= 1.CIOOE-C14
optimization parameter 1 ...... .. ..ctptl= 1.000E+OO
optimi=atlon parameter 2 ........ ..ctpt2= O*OCIOE-01
ctptimi:atlon parameter 3 . . . . . . . . ..apt3= C}.CIOOE-01

availabl~ prc~babilit~ distribtian%:

nc,rmal ... . . . . . . . . . .........ndis=l
lcognoriiial................ndis=~
gamma i. ............. .....ndis=3
exponential ..............ndis=4
ra,yleigh .......... ........nd is=5
uniform ...................ndis=6
beta ... ...................ndis=7
type i largest value .....ndis=ll
type i smallest value ....ndis=12
type ii largest value . . ..ndis=13
Welbull.. .. . . . . . . . . . . . . . . .ndis=14

statistical data cif basic varibl~s:

var ndis m=an st. dev. paraml pararn2 param3 param4 init. pt
1 11 1.00E+02 2.IXE+CI1 9.lCIE+O1 6.41E-02 1.6CIE+<}2
2 2 2.00E+O1 5.00E+OCI 2.97E+O0 2.46E-01 2.OC}E+O1

draterministic par~meters in performance function:

dp t 1) = 1.00CIE+OO
dp ( 2) = !50000E-01

**** salutian phase *+ y.*

mvfclsm technique: beta = 0. 9806; failure probability = 1.634E-01

---------------------------- --------------------
iteratic~n na. 1
var. lirtearisatien pctint unit normal
alpha
1 1.CIO;E+C12 1.77gE-C)l c1.lW%
2 2.OCICIE+O1 1.231E-01 -0.9817
reliability index beta = 0.2159
---------------------- --------------- --._----_-
iteraticm no. 2
var. linearization paint unit n~rmal
x alpha
1 9.992E+01 1.731E-01 c1.3046
2 1.!557E+01 -8.93QE-01 -cl.9s25 .
reliability index beta = 0.9097
-----------------------
.. ---------------,---.,-- ---
iteraticlmna. 3
var. #ineari=atian point unit normal
.
alpha
1 1.04;E+02 3.78;E-Q1 0. 3&97
2 1.449E+01 -1.1S5E+O0 -0.9292
reliability index beta = 1.2437
-+.---- ----- . -.,-..

327
 iteratic,n nc~. 4
var. lineari2atic1n paint unit narmal
x Y alpha
L 1.038E+02 4.636E-01 c1.3777
2 1.4E&E+01 -1.17CJE+C10 -CI.9239
reliability index beta = 1. 259!5
------------------------- -------------------- --
iteratictn no. 5
var. linearization unit nurmal paint ..
Y alpha
.
~~.c)A;E+02 4.7SSE-01 0.3781
: &4S6E+01 -1.166E+O0 -CI.9258 +
reliabil~~ index beta = 1. 2S89
----------------------- ------------------------- -------------------------- ---
iteratic~n no. &
var design puint unit sensitivity vectsrs nearest pt unit ficmmal
X* gamma delta eta y* alpha
1 lo040E+02 0.3702 -i~.271s -().C)841 4.761E-01 0.3782
2 1.4S&E+01 -cl.9257 c).9624 -0. ?965 -1.165E+CJCl -cl.9257
----------------------- ---------------------- -------------------------- ------
reliability index beta = 1. 23g9
failu~e probability = 1.04CIE-01

***************+*****************+********************
*
Sensitivity Analysis
* *
***************+*****************+********************

+* Sensitivity Analysis an Distrik,uticm Parameters **

d(beta)/d(parameter) :
var mean ~td dev par 1 par 2 par 3 par 4
1 -1.773E-C12 -5.349E-03 -1.773E-02 4.157E+CM3
2 2.514E-01 -2.536E-01 3.7SJE+CKI -4.3S2E+CKI

d(pfl)fd (parameter) :
var mean !atd dev par 1 par ~ par 3 par 4
1 3.2C12E-03 9.661E-04 3.202E-CJ3 -7.5C17E-01
2 -4.54CIE-02 4.58CIE-02 -6.791E-01 7.914E-CI1

w++ sensitivity Analysis an Deterministic Farameter6 W+W

par d(beta)ld(parameter) d(pfl)id(parameter )

1 -1.880E+O0 3.393E-01
2 3.760E+Ci0 -h.?91E-01
-.
&
:Y*

. . . .. .
*---

328
 ********+*********************************************
* *
* Second Order Structural Reliability Analysis *
* *
Point Fitting lleth~d
* *
******************************************************
. .

:.+
++ coordinates
-1
? and ave. main curvature= cif fitting pctints im ratated space +*+
+
Y 1 = ;.259944 yz= 1.216377
Y*1= -1.250944 y2= 1.e1750b
a 1 =-c}.2&47&lSE-01
------------------------- -.-_---&------------- -----_--_---------

*** s.ecand-arder appraximatian ***

failure generalized
probability reliability index
Breitung a=ymptcltic fc~rmula 1.077E-01 1.2390
Tvedt tbre~ term fclrmula 1.091E-01 1.2315
Iwedt single integral formula l.CW.IE-01 1.2316
Tvedt dcmble int~gral formula 1.090E-01 1.2316

**********+***************+ ***************************
*
Secand Order Structural Reliability Gnalysis *
* #
* Curvature Fitting Method *
*
*****+*******+**+*************************************

+** curvature matrix at dte~lgn pc,int in rc.tat~d space **+

1
1 -2.7WE-02

secclnd-arder apprnximatian **

failure generalized
probability reliability index
Breitung a5ymptGtic fctrmula . 1.078E-01 1.2303
Tvedt three term fc,rmula 1.092E-01 1.2306
Tve-dt single integral formula 1.C)92E-01 1.2306
Tvedt dcluble
.7= integral formula 1.092E-01 1.2306
~ :b
s
-- ~
a.

. . - - -.. -

329
 4

**+******+* ****************4************ *******************

* Un i ver s i ty a f Cd 1 i fern i a *
* Department OfCivil Engineering *
* Divisian of *
* Structural Engineering and Structural Mechanic=. *
* *
*. F OSR A P *
* First Order Structural Reliability Analysis PrQgr~m
-.
* *
i. Devel~ped By *
Pei-Ling Liu and Armen Der Kiureghian *
Last Revision: September 1986 *
* *
* Extend=d for FOFD and SOSRAP by *
* HONG-ZONG LIN, March 1986 *
**+******************************************** *************

**** input data -- prclblem 1 ****

examplE ~
type af first-order technique used .. .....ifo= 3
iffi=l ...rnean-valueq 1st carder, 2nd mcunent (mvfosm) methctd
ifc,=z .~.firgt-c{rder, secc,nd mament (fosm) method
ifo=3 ...first-cjrder~ marginal distribution (fomd) method
ifc,=4 . ..fir~t-order$ full dist. (Rc,senblatt trans) methed
type clf ~ecc~nd-order technique used .....iso= 3
i%a=(} ....... ......no second-arder appreximatien
isci=l . .... =.. ... ... ........paint fitting methed
iso=2 ....- .. ...........curvature fitting methed
ise=3 . . . . . ..paint and curvature fitting m=thed~
type c,f integrdtien schemes u~~d in seccmd-firder
analysis ..... .. . ..... .. . ......*1tg=
1

itg.eq.o ....Ereitungs Tvedts 3-term and single integral

itg.ne.Cl .. all af the atc~ve plus Tvedts double i~tegral
type clf sensitivity analysis required ...isv= 3
isv=CJ ........ ...... ......no sensitivity analysis required
i SV= 1 . . . . . . . . . . . . .sensitivity
of distribution parameters
isv=2 .......sensitivity c{f perfcomdnce function parameters
i~v=3 . ...distributian and performance function parameters
nufiiberof randc,m variables .... .. ... ... ...nx= 2
number c,f deterministic parameters .... ...np= ~
number c,f user prc,vided dl~tributlctn . . . . . mu= o
numh~r clf parameter= in user-defined full
distributions .. ..... ..... ... ......... ....ns= o
cc,rr~latictn structure . . .. == =..=.. =.ncarr=rr= 1
ncclrr.~.CI . . . . . . . . . . . . . ..uncctrrelated variable=
ncc,rr.~.0 ... ............ ..correlated variables .-
initiali~tien flag ........ ...... .......inlt= 0 T
g
init.eq;O ..........initialixatisn at mean paint r..
init.ne.O .... ... ... ......initiali~ation by u=mr r:
@utput flag . . . . . . . . . . . . . . . . a . . . . . . . . ipr= . 1
lpr.~q.0 . . . . . . . . . . . . . . Output anly final results
lpr.ne.O . . . . . . . . . . . . . .Clutput at every ipr St=ps
--- -+_-

330
 aptimizatic~n ~cheme used . . . . . . . . . . . . . ..lCipt= 1 4
icapt=l ............................ .RF-HL method
ictpt=~ .......... ..... .....Mcldified RF-HL mettmd
, ic,pt=3 . . . . . . . . . . . . . ..Gredient %ojecti~n method
maximum number Of iteratiun cycl~s . . . ..nitl= 20
ma>.lmurn =teps In line search ... .. ... .. .ni t2= 4
cctnvergsnc~ tal=rance . . . . . . . . . . . . ..tc.l= 1.000E-C)4
c~ptimizatlcm parameter 1 ..........optl= 1.000E+OO
Ciptimization parameter 2 ....... ...c*pt2= O.OCKIE-01
c~ptimizaticm parameter 3 . . . . ... . ..c.pt3= 0.000E-01
.-_:
availabl~ prabatility distributions:
,=
normal,
. .... ..............ndis=l
lngnarmal . . . . . . . . . . . . . . ..ndis=2
gamma . . . . . . . . . . . . . . . . . . ..ndi~=q
e::panential . . . . . . . . . . . . ..ndis=b
rayleigh . . . . . . . . . . . . . . . . ..n d is=S
unlfctrm . . . . . . . . . . . . . . . . . . .ndis=6
beta .......... ...... ......ndis=7
type i largest value .....ndis=ll
tfp~ i =mallest value ....ndiS=12
type ii lar~est value .. ..ndis=13
welbull . ..................ndis=14

statistical data clf basic varibles:

var ndis m-an St . dev. paraml param2 param3 paratri4 init. pt
1 11 1.00E+C12 ~.~ctE+~l 9.ICIE+O1 6.41E-02 1.<~CjE+~2
2 2 2.C@E+Ol 5.QOE+OO 2.97E+C)0 2.4&E-C)l 2.00E+C)l
.
d~terministic parameters in performance functian:
dp ( 1) = 1.CIC)OE+OO
dp ( 2) = 5.000E-01

c~rrelaticm coefficient matrix in Original space,:

12
1 1 CKJ
f3.5cl
2 0.50 1.00

.carrelaticln cc~efficient matrix in narmal space:

1
1 1.00 :.51
2 0.!51 1.C)o

. .- ~.
**** suluticm phase ****

wfc,sm technique: beta = 1.6911; failure probability = 1.376E-Cll

------------------------- ----------------, -------
iteration ,rmc. 1 ...-----
~ar . lineari~aticm pOint unit normal
,.
:- x alpha
1 ~iCICKtE+02 1.77;E-01 -0.3470
2 ~; OCIOE+Ol 3.764E-02 --0.9379
eliabilty index beta = 0.1813
--------------------- ------------- -------------
:teratian rm. 2
far. lineari=aticin pc~int unit nermal
x lpha
1 9.C~78E+Cl -3.51[E-01 -0.2407
z 1.519E+C~l -9.489E-01 -0. 9b54
~liabilty index beta = 1.0117
~-. .. . .,,.
331 --
 pa+ d(beta)/d(pararrt~ter ) d(pfl)/d(parameter )
it -2.32SE+OC) 2.7azE-cll
2 4.6ZAE+CK) -S.E&3E-C11

*********~+**+**+*******++*****+4****+****************

~~1=
1.552342 Y2= 1.4&3146
y?l= -1 .!552B42 Y2= 1.43.5239
a i =--lJ.3d4sga3E-cll
------------------ ________ ______ ___ ______ _______ ______ ___

.-

+** curv~~ure matrix at design p~int in rotated .pace **+

1
1 -2.27~E-C)2
..
.,:
--
.=. ;-.
~_
., *** secclnd-order apprc{ximatian *** ::

fai lure generalized

prcibability reliability j.ndex
Ereltlmg asymptc,tic fc,rmula &.3E&E-02 1.5257
lvedt three term fclrmula &.446E-C12 1.51s4
T~~dt ~ingle int~gral fclrmula 6.44EIE-C@ 1.5195
Tvedt dcubl~ integral fc!rmula 6.445E-C12 1.5184

332
 function g(x~dp)
implicit real*S (a-hqa-~)
1 dimensicln x(l) !dp(l)
9 = alp(l) - dp(2)*x(l)
return
- dp(3)*x(2) ~ = 5-x, -xz
end

.
subrciutine df(par~x!nd~cdf?pdf)
implicit real*8 (a-hza-z)
difi~nsiun par(4)
rekrn
end,

subrcmtin~ hfun(x,ih?d~th)
implicit real+S(a-h9e-z)
dimensien x(l),ds(l)
ga to (10!2C1) ih
10h= l-dexp(-x(l))
r~turn
2~h= l-( l+ds(l)+x (s))+de:~p(-x (2)-d~( l)*~(l)*::(2))
return
end

233
 +****+*******+***+*****+**s.**********+*******+**+***********
+ Un i v~r s i t y c, f Ca 1 i for n i a *
* Department cl
f Eivll Engineering *
* Division clf
+ Structural Engineering and Structural Mechanics *
*
* CALREL
* First and Second ~rder Reliability Analysis Prc~gram *
*
- *
*,.> Develc.p~d By . *
* Pei-Ling Liu and Armen Der Kiureghian *
*. *
Last Revision: September 1986
* *
* Extended fcrr F!3FD and SOSRAP by *
* HONG-ZDNG LIh!r Ilarch 19S6 *
**+**+*******+* ***+***************+*+***+*****++***+**++****

example 3
type cIf first-order tschnique used .. .....ifo= 4
i%c=l ...cIeari-value~ Ist cir-derz 2nd mclment (mvfclsm) m~tticid
ifc=2 . . .fir~t-c!rd~r$ s=cc~nd Mcim=nt (fosal) ,methctd
if~=3 ...fir~t-c~:-der~ marginal distributic!n (fc~ma) methctd
i~c~=b . . .flrst-cli-der> fuil dist. (Rcts~rhlatt tran=) methctd
type Cif SFcct.-lti -c,i-d~r tgchnique used ... ..isc= 3
isc~=O ............ .nc! secclnd-or~sr aFprciximatlan
ism=l .. . . . . . . . . . . . . . . . . . . ..paint fitting methctd
iscl=~ . . . . . . . . . . . . . . . . ..curvatur= fitting method
i=cl=~ . . . . . ..pc~lnt and curvature fitting metheds
type of ir,tegratic:n schemes used in secc!nd-~rder
dnalysls . ....*.. ... .. .. ....... . ..... .-. ltg= 1
]tg.r.q.(} . . . .H.-eitL~rlgsTvedtss 3-term and slnqle integral
i,tg.rl-.0 . . ll clf the abcrve plus Tv~dt~ C!cluble integral
type Clf %Snsitivit y ana.ly~i~. req~.;ired ...isv= 3
i~.V,=<) .. ........ . .. .....nci serlsitivity analysis r~quirec!

ls~=l .. ..... .... ...sensl tivity ctf di~tributicln parametei-s

i~v=a ... . . . .s~n~itlvity of perf:imance functimn parameter~
i~v=s ....di%trihutic~n and perfc,rmance functic{n parameters
numbsr clf rahdcim variables .... .. .........nx= 2
number cif deterministic parameters .......np= 3
number clf q==~< prcivided distributicsn .....nu= Q
nl.ltttl~i- CI+ psr;meter~ in user-defined full
diiGtrlbLltiC!l15 . . . . . . .. . . . . . . . . . . . . . .. .....ns= 1
cctrrelaticliastructure .. ..... . .. . . .....ncorr= 1
ncfirr.~q.C~ .... . .. .. ......unc~rrelat~d variables
rtcorr.ne.~) .. .. .. .... ... ... .correlated variable%
li-latl~l,i:-atlc,n flag .. . . . . . . . . . . . . . . . . . ..in]t= 1
ir,it.eq+;t} . . . . . . ... .i.nitializaticm
at mean pc~int
. , .. . .
,lnxt.n~ ~~ ..... ....... ....init]ali=atic~n by user
Output flag .... . .. -,......... . . ........ipr=
1
ipr.,eq.(1 ..... .. .. .. .. .c,utput only final results
ipr.r,e.Ci . . . . . .. i . . . ...cutput at every ipr steps

334
..
.
 ic,pt=l ......... ............... . .. ..RF-HL methcld
ic,pt=2 . . . . . . . . . . . . . . . . . . . .flc,dified RF-HL methc,d
sc!pt=3 . . . . . . . . . . . . . ..(%adient Prajecticin method
q!imum number c[f iters,ticln cycles . . . ..nitl= 20
rna::imum steps in lirle search --.*... . ...ni t2= 4
ccnvergemce tcrler=nce . . . . . . . . . . . . . tc,l= 1.00CIE-134
.
c+ptlml:atlc~n parameter 1 . . . . . . . . . . c~ptl= 1. [:@t:~E+C)~)
ctptimizaticln p~ram=t~r 2 . . . . . . . . . . clpt~= fJ.~c~~)E-01
c,ptirr,izatic,n param~ter 3 . . . . . . . . ..r~pt3= O.OW)E-01

------------------------- ----------------------
iteraticln no. 1
var . linei~rizatictn paint unit nc~rmal
>: alpha
1 1.82SE-C)$ -9.&E-CJl -0.42!54
2 4.R19E+CK~ z.t)b4E+~cI c1.9135CI
r~liahjlty index keta = 2. 27s8

it~raticln rlcl. 2
var. linearizatic,n pc,int unit nc,rmal
.~ Y alpha
1 1.S17E-01 -9.65E-Cll -(2.4246
2 4.S18E+OC1 2.{}A2E+O(} Cl.9254
r~liabi.lty ind~x beta = 2. 2?E@
--------------------- ------------------------ --
iteraticin nci. 3
var. linearizaticin pctint unit nclr.mal
.A alpha
Y
1 1.823E-01 -9.475E-<11 -0. 42!52
2 4.G15E+O0 Z.(:163E+C)0 CI.9CE1
reliability index bcita = 2. Z78E!

iteraticlr m. 4
Var . linearizaticlri pclint unit normal
X Y alpha
i 1.SISE-01 -9.690E-c}I -CJ .4247
E 4.3:SE+O0 ~.(:)L.zE+{:)<i 0.?0!53
reliabzlty index beta = 2 27a8
------------------------- ----------------------
iteraticln ncl. 5
var . linearization pclint unit nurmal
:.: Y alpha
1 1.S%2E-01 -9.679E-111 -CI.42Eil
2 4.91BE+CKI 2.cI&3E+C)CI 0.9<)51
reliability inde:: b@*a = 2.2700
---------------------- -------------------------
iteratictn nc,. &
var. lirlearlzaticln pclint unit nc!r.mal
-- alpha
Y .
1 ti;91;E-1:)1 -9.68?E-01 -0.424S *.
2 4.81BE+OCI 2.C)63E+CM3 ().9()53
r~liabilty inde:: beta = 2. 27S8
-------=- --------------------- ----------------- . -. .
itfars.tic~n m,. 7
var . llrl~arizatinn
pciint unit nclrmal
X Y alpha
1 1.E)21E-01 -9.A91E-(:J1 -{1.42Z0
2 4.PISE+CWJ 2.C16SE+C}0 0. 90S2
reliability index beta = 2.2708
---- - . ----------- -------
--?
 var. L1l-1e-1 liaLAL*Il p~.~i,+ -. .-.,.+ ,. -----
4
alpha

x

1 1. 82 C)E-O 1 -9 68[E-01 -0.4249

0.9052
re~i abi 1~~=;%~~0beta2 C)S3E+** Z .2708
--------------------- ------- ------- ------- .
iteraticln no. 9
var. linear izaticln pctint unit nc~rmal
X alpha
1 43;E-C)l
1 .S21E-CI1 -Cl 4250 -9.

2 4. HlSE+@3 2. 063E+OC) 0. 9C152

reliabilfy index beta = 2.2708
..--
------- -e---- ------------------------ --------- i-
iteration.no. 10
var. lin~arizatictn point unit normal
x Y alpha
1 1.82C1E-01 -9.h85E-01 -0.4249
2 4.81SE+C@ 2.C163E+C@ 0. 9Cl!52
reliability index beta = 2. 27%0
--------------------- --------------- _______ ----
it~ratic,n nc~. 11
var. lirt~arizatic,n pc,int unit nclrmal
x Y alpha
1 -9.&S3E-01
1.S21E-C~l -0.4250
2 4.SISE+CK1 2.CM3E+C@ 0.9052
reliability index beta = 2. 273G
-------------------------- ------ _____
iteratican no. 12
var. lineari:aticm pctint unit normal
x alpha
1 1.~2c}E-~1 -%68:E-01 -0.4249
2 4.=18E+WI 2.063E+C)0 0.9052
reliability index beta = 2.2700
---------------------- --------- ______
reliability index beta = 2- 278e
failure prctbability = 1.134E-C)2

******************+*************************+*********
* *
* Sensitivity Analysis *

**+**************+************************+*******+***

-*** Sensitivity Gnalysis en Distributing Parameters ***

par d(beta)/d(parameter) d(pfl)/d(parameter )

1 1.SZ5E-C}2 -S.427E-04
:...
...:
=
~T ** Sensitivity Analysis on Deterministic Parameters ***
s i:
par d(beta)/d(parameter) d(pfl)/d(parameter) .
1 3.76EIE-CI1 -1.l19E-02 .:
2 -6.833E-02 2.039E-03
3 -1.814E+C10 5.393E-C12

.. . . .

336
~, j = 2.273824 ~.. ~ = 2.545[:)3<)
~s~= -Z.27W324 ~F~= 2.5%9305
s 1 = ~).=f:JL@2&pE-~~l
------ ------ ----- ----- ------ ----- ----------- _____ ----- _____ _____

fai lure g=neraliz=cl

prc~bability - r=.liability ind~x
13!-eiturlg
a~ymF?,cttic farmula 9.5:2E-03 2.343?
T\vedt,three ~~im fc,rmula 9,295E_~:)3 2.3541
Tv@fit =irqle integral farmula 9.2?IE-<13 2 3530

Tve5t cic,uttleinteqral fcrmula 9.5+SE-(113 2.353B

337
 Ta run ChLREL, da the following:

(1) Turn cin the pawer of IBtlPC and VCR terminal.

(2) Create a temparary directary in.drive C and make it yaur

b.
working directory.
tC:\l> MD CE249
. ;
g CC:\l> CD CE249 s
,> +
- (3) Copy all files in D:\CE249 to drive C.
.
CC:\CE2491> COPY l):\CE24q\*.* .

(4) Reset the search path.

tC:\CE2491> START

(5) Edit the three subrcwtines in u==r.far SUCh that they work
in the same manner as indicated in the cla=s handaut.
Tcc edit the subroutines, key in:
CC:\CE2493> FE USER.FOR
TO display the help file in PEJ press <Fl> key.

(6) Cctmpile user.f~r nd link the prc~gram.

CC:JCE2491> CUSER
If there are errors, use *TYPE ERR ta examine theml and
cclrrect the errors in user.fer. Then compile user.far again.
Once a c~rrect user.far is ctimpiled~ link the pragram.
tC:\CE2491> LREL

(7) Create am input data file.

tC:\CEZ491> PE IN
Ta exit the editnr, type <F3> IN NOTAEIS <RETURN>. If NOTAiis
is skippedl same spaces in the input file will be replaced
by tabs. ThatG,mes&es up the input file.

Run pregram C~LREL.

CC:\CE2493> CALREL < IN (Output an screen)
tC:,\CE249J> CALREL < IN > OUT (cwtput te file OUT)
Make sure you leave spaces between file names and the SYmbc,l.s .
\ and >.

(9) Read c,utput file (if the output iS raut~d &a a file)
KC:\CE2491} TYPE OUT (read the file OUT on
screen)
tC:\CE2491> NETFRINT OUT (get a hardeapy of the output file)
Befclre the print queue is mbmitted~ make sure that the line
printer is enline.
..-. . , .. ....... . .
(10) Stare yaur files in a flctppy disk and delete ll files in
-.
...
,..
. the working direet~ry.
*.=
. [C:\CE249J> COPY (yaur file) A:
C
G
~- EC:\CE2491> ERASE *.* ~
&
A
(1$) Delete the temporary directery. *
CC:\CE2493> CD \
tC:\J> RD CE249

[12) la shut dawn the machine, take yaur disk~tt~ Gut and turn

off the power.

 COMMITTEE ON IWRINE STRUCTURES

Commission on Engineering and Technical Systems

National Academy of Sciences - National Research Council

The COMMITTEE ON MARINE STRUCTURES has technical cognizance over

the interagency Ship Structure Committees research program.

Stanley G. Stiansen (Chairman), Riverhead, NY

Mark Y. Berman, Amoco Production Company, Tulsa, OK
Peter A. Gale, Webb Institute of Naval Architecture, Glen Cove, NY
Rolf D. Glasfeld, General Dynamics Corporation, Groton, CT
William H. Hartt, Florida Atlantic University, Boca Raton, FL
Paul H. Wirsching, University of Arizona, Tucson, AZ
Alexander B. Stavo~, National Research Council, Washington, DC
Michael K. Parmelee, Secretary,- Ship Structure Committee,
Washington, DC

LOADS WORK GROUP

Paul H. Wirsching (Chairman), University of Arizona, Tucson, AZ

Subrata K. Chakrabarti, Chicago Bridge and Iron Company, Plainfield, IL
Keith D. Hjelmstad, University of Illinois, Urbana, IL
Hsien Yun Jan, Martech Incorporated, Neshanic Station, NJ
Jack Y. K. Lou, Texas A & M University, College Station, TX
Naresh Maniar, M. Rosenblatt & Son, Incorporated, New york, NY
SOIOmOn C. S. Yim, Oregon State University, CoHallis, OR

MATERIALS WORK GROUP

William H. Hartt (Chairman), Florida Atlantic University, Boca Raton, FL

Fereshteh Ebrahimi, University of Florida, Gainesville, FL
Santiago Ibarra, Jr., Amoco Corporation, Naperville, IL
Paul A. Lagace, Massachusetts Institute of Technology, Cambridge, MA
John Landes, University of Tennessee, Knoxville, TN
Mamdouh M. Salama, Conoco Incorporated, Ponca City, OK
James M. Sawhill, Jr., Newport News Shipbuilding, Newport News, VA

...

C3
 SHIP STRUCTURE COMMITTEE PUBLICATIONS

SSC-334 Influence of Weld Porosity on the Inteurity of Marine

Structures by william J. Walsh , Brian N. LeiS, and J. y.
Yung 1989

SSC-335 Performance of Underwater Weldments by R. J. Dexter, E. B.

Norris, W. R. Schick, and P. D. Watson 1986

SSC-336 Liquid Slosh Loadina in Slack Ship Tanks; Forces on

Internal Structures & Pressures by N. A. Hamlin 1986

SSC-337 Part 1 - Shi~ Fracture Mechanisms Investigation by Karl

A. Stambaugh and William A. Wood 1987

SSC-337 Part 2 - Ship Fracture Mechanisms - A Non-Experts Guide

for Inspecting and Determining the Causes of Significant
$hi~ Fractures by Karl A. Stambaugh and William A. Wood
1987

SSC-338 Fatiaue Prediction Analysis Validation from SL-7 Hatch

Corner Strain Data by Jen-Wen Chiou and Yung-Kuang Chen
1985

SSC-339 Ice Loads and ShiD Response to Ice - A Second Season by

C. Daley, J. W. St. John, R. Brown, J. Meyer, and I. Glen
1990

SSC-340 Ice Forces and Shi~ Response to Ice - Consolidation.. REp.ort

by C. Daley, J. W. St. Johnr R. Brown, and I. Glen 1990

SSC-341 Global Ice Forces and ShiD ResDanse to Ice by P. Minnick,

J. W. St. John, B. Cowper, and M. Edgecomb 1990

SSC-342 Global Ice Forces and Shi~ ResDonse to Ice.: Analy~is of

Ice Ramminu Forces by Yung-Kuang Chen, Alfred L. Tunik,
and Albert P-Y Chen 1990

SSC-343 Global Ice Forces and Ship Response to Ice - A Second

Season by p. Minnick and J. W. St. John 1990

SSC-346 Fatiaue Characterization of Fabricated Ship Details -

Phase 2 by K. K. Park and F. V. Lawrence, Jr. 1988

SSC-349 Develo~ment of a Generalized Onboard Res~onse Monitoring

System (Phase 1) by F. W. DeBord, Jr. and B. Hennessy
1987

SSC-350 ShiD Vibration Desiqn Guide by Edward F. Noonan 1989

None Ship Structure_~ommittee Publi~tions - A Special

Bibliography 1983