Sei sulla pagina 1di 263

Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.

866494
Purchased from American Institute of Aeronautics and Astronautics

Some Engineering Applications


in Random Vibrations
and Random Structures
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

Some Engineering Applications


in Random Vibrations
and Random Structures
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Giora Maymon
RAFAEL
Haifa, Israel

Volume 178
PROGRESS IN
ASTRONAUTICS AND AERONAUTICS

Paul Zarchan, Editor-in-Chief


Charles Stark Draper Laboratory, Inc.
Cambridge, Massachusetts

Published by the
American Institute of Aeronautics and Astronautics, Inc.
1801 Alexander Bell Drive, Reston, Virginia 20191-4344
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Copyright O 1998 by the American Institute of Aeronautics and Astronautics, Inc. Printed in the United
States of America. All rights reserved. Reproduction or @anslationof any part of this work beyond that
permitted by Sections 107 and 108 of the U.S. Copyright Law without the permission of the copyright
owner is unlawful. The code following this statement indicates the copyright owner's consent that copies
of articles in this volume may be made for personal or internal use, on condition that the copier pay the
per-copy fee ($2.00) plus the per-page fee ($0.50) through the Copyright Clearance Center, Inc., 222
Rosewood Drive, Danvers, Massachusetts 01923. This consent does not extend to other kinds of copying,
for which permission requests should be addressed to the publisher. Users should employ the following
code when reporting copying from the volume to the Coypright Clearance Center:

Data and information appearing in this book are for informational purposes only. AIAA is not responsible
for any injury or damage resulting from use or reliance, nor does AIAA warrant that use or reliance will be
free from privately owned rights.

ISBN 1-56347-258-9
Purchased from American Institute of Aeronautics and Astronautics

Progress in Astronautics and Aeronautics


Editor-in-Chief
Paul Zarchan
Charles Stark Draper Laboratory, Inc.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Editorial Board
Richard G. Bradley Leroy S. Fletcher
Lockheed Martin Fort Worth Company Texas A&M University

William Brandon Allen E. Fuhs


MITRE Corporation Camel, California

Clarence B. Cohen Ira D. Jacobson


Redondo Beach, California Embry-Riddle Aeronautical University

Luigi De Luca John L. Junkins


Politechnico di Milano, Italy Texas ABM University

Philip D. Hattis Pradip M. Sagdeo


Charles Stark Draper Laboratory, lnc. University of Michigan

Vigor Yang
Pennsylvania State University
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

Foreword

The uncertainties that occur in the design process, in the models employed
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

for analysis of the loads, and in the geometric parameters of a structure,


have been dealt with for generations by experience and safety factors. In
recent years, however, the deterministic approach has been enhanced, and
in some cases even replaced, by a probabilistic approach in which the
various uncertainties are treated more rigorously.
These methods of design and analysis require the analyst and designer
to have a better understanding of random processes, random vibrations,
and random structures. In this book, the author tries to provide this knowl-
edge and brings together a careful introduction to the fundamental theory
with many different practical techniques for analysis of random vibrations
in structures.
Though it is assumed that the reader has a rudimentary knowledge of
random vibration theory and probability theory, the basic concepts and
definitions are outlined (with the relevant references) and only then em-
ployed in the development of practical analysis techniques. The presenta-
tion therefore addresses the needs of both experienced design engineers
and newcomers to the field.
The emphasis is on engineering applications, some of which are presented
in considerable detail, including flowcharts for numerical procedures. The
matching with commercially available, large finite element programs is also
extensively discussed, as well as the capabilities of commercially available
special programs for the analysis of probabilistic structures. Failure analyses
of dynamic systems and models for crack growth, both deterministic and
stochastic, are treated, accentuating the advantages of the stochastic models.
The focus is on engineering applications, as stated in the title. Based on
his extensive experience as a designer and researcher, Giora Maymon
presents the material in a manner appealing to the design-oriented engineer,
while preserving the necessary sound mathematical basis. It is hoped that
the book will familiarize the engineering community with random vibrations
and with stochastic methods and will encourage engineers to study and
adopt them widely in design procedures.
Josef Singer
Professor Emeritus
Technion-Israel Institute of Technology

vii
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

Table of Contents

..................................
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Preface xiii

Chapter 1 Deterministic Single-Degree-of-Freedom System .. 1


I. SDOF System Subjected to External Harmonic Excitation . . . .
I1. SDOF System Subjected to Base Excitation ............
111. Response of an SDOF System to General Force . . . . . . . . . .
IV . Stresses in an SDOF System .....................
V. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 2 Deterministic Multiple-Degree-of-Freedom System .


I. Differential Equations and Normal Modes of an MDOF System
I1. Generalized Masses. Dampings. Rigidities. and Forces ......
111. Uncoupled Differential Equations . . . . . . . . . . . . . . . . . .
IV . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 3 Deterministic Continuous System ...........


I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I1. Differential Equations of Continuous System . . . . . . . . . . . .
111. Base Excitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IV . Stress Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
V. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 4 Random Functions and Excitation ...........


I. Basic Concepts of Random Functions . . . . . . . . . . . . . . . .
I1. Practical Characterization of Random Excitation . . . . . . . . . .
I11. Important Excitation Functions . . . . . . . . . . . . . . . . . . . .
IV . Boundary-Layer Excitation Model . . . . . . . . . . . . . . . . . .
V. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 5 Response of Linear Systems to Stationary Random


Excitations .................................
I . Response of a Linear SDOF System . . . . . . . . . . . . . . . . .
I1. Response of a Linear MDOF System . . . . . . . . . . . . . . . .
Purchased from American Institute of Aeronautics and Astronautics

x G . MAYMON

I11. Response of a Linear Structure to Stationary Random


Excitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IV . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 6 Nonlinear Single-Degree-of-Freedom and


Multiple-Degree-of-Freedom Systems ................
I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I1. Nonlinear Behavior of an SDOF Oscillator . . . . . . . . . . . . .


I11. Nonlinear Coefficients of a Structure ................
IV . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 7 Statistical Linearization Method ............


I . Statistical Linearization Method for an MDOF System . . . . . .
I1. Nonlinear Response of an SDOF System to Random Gaussian
Force ..................................
I11. Nonlinear Random Response of Two-DOF System to Random
Gaussian Force . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IV. Nonlinear Random Response of an Elastic System . . . . . . . .
V . Computational Procedure . . . . . . . . . . . . . . . . . . . . . . .
VI . Calculation of Stress Response . . . . . . . . . . . . . . . . . . . .
VII . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 8 Nondeterministic Structures: Basic Concepts .....


I. Introduction . . . . . . . . . . . . ...................
I1. Failure Surface: Basic Case . . . ...................
I11. Reliability Index ......... ...................
IV . Summary . . . . . . . . . . . . . . ...................

Chapter 9 Calculation of the Probability of Failure .......


I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I1. Lagrange Multiplier Method .....................
I11. Demonstration of the Iterative Process ...............
IV . Numerical Programs for Probabilistic Structural Analysis . . . . .
V. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 10 Taylor Series Expansion of the Failure Surface ...


I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I1. Taylor Series Expansion . . . . . . . . . . . . . . . . . . . . . . . .
I11. Selection of the Evaluation Point . . . . . . . . . . . . . . . . . . .
IV . Detailed Examples of the Taylor Series Expansion Method ...
V. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Purchased from American Institute of Aeronautics and Astronautics

CONTENTS xi

Chapter 11 Direct Calculation of the Probability of Failure Using


an Existing Finite Element Program ................. 157
I. Introduction . . . . . . . . ....................... 157
I1. MJPDFMethod . . . . . . ....................... 158
111. Numerical Examples . . . ....................... 163
IV . Summary . . . . . . . . . . ....................... 178

Chapter 12 Probability of Failure of Dynamic Systems ..... 179


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
I1. Statistical Behavior of a Stationary Gaussian Process . . . . . . . 179
111. Spectral Moments of a Determinstic MDOF System . . . . . . . . 184
IV . Probability of Threshold Crossing (Failure) . . . . . . . . . . . . . 185
V. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

Chapter 13 Stochastic Crack Growth Models ........... 195


I. Introduction .............................. 195
I1. Stochastic Crack Growth Models .................. 196
111. Crack Length Distribution ...................... 202
IV . Failure Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
V. Probability of Failure Calculation . . . . . . . . . . . . . . . . . . 212
VI . Required Information for the Analysis . . . . . . . . . . . . . . . 215
VII . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217

Chapter 14 Concluding Remarks .................. 219

Appendix A: Some Important Integrals .............. 221


I . Constant Value PSD Functions . . . . . . . . . . . . . . . . . . . . 221
I1. Filtered One-sided White Noise .................... 222

Appendix B: Conversion Between Acoustic Decibels and PSD

Appendix C: Finite Element Input Files for MJDPF Method


1. File for Example 1 . . . . . . . . . . . . ...............
I1. File for Example 2 . . . . . . . . . . . . ...............
I11. File for Example 3 . . . . . . . . . . . . ...............
IV . File for Example 4 . . . . . . . . . . . . ...............
V. File for Example 5 . . . . . . . . . . . . ...............
VI . File for Example 6 . . . . . . . . . . . . . . . . . . . . . . . . . . .

References ................................
Index ................................... 245
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

Preface
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The purpose of this book is to present design engineers with some of the
basic tools for the solution of practical problems they may encounter when
analyzing random vibration of deterministic and random structures. The
content of this book is presented so that it enables them practical use of
the procedures and expressions required for daily engineering work. Having
more than 35 years of experience in structural dynamics research and
development, the author believes that a good understanding of the basic
concepts and the physical meaning of the behavior of a structure is crucial
to a successful work of engineers, especially in the beginning of their careers,
and the material presented in this book is directed toward this goal. Never-
theless, in many practical cases, numerical solutions are the only way to solve
some of the practical cases, so flowcharts for such numerical procedures are
presented for many of the discussed cases. Using these flowcharts, each
user can take advantage of a favorite mathematical solver, i.e., MATLAB,'"
TK Solver, and others, or write a Fortran program to build a numerical
procedure suitable for the design needs.
The design of any engineering system is a process of decision making,
under constraints of uncertainty. The uncertainty in the design process
result from the lack of deterministic knowledge of different physical param-
eters and the uncertainty in the models with which the design is performed.
This is true of many of the disciplines involved in any design such as
electronics, mechanics, aerodynamics, and structures.
The main sources of uncertainties in structural analysis and design are
1) uncertainties in the determination of the physical and mathematical
model used for the analysis, including uncertainties in the failure criteria
(model); 2) uncertainties in the determination of the magnitudes, locations,
frequency content and correlations of the external (static and dynamic)
loads (load); and 3) uncertainties in various structural parameters such
as geometries, dimensions, material properties, and allowables (stochastic
structure). These three categories do not include other more subjective
uncertainties such as human errors in the design and production.
Uncertainties in the model were treated by the scientific and engineering
communities by exploring in three major directions:
1) The first uses analytical models for linear and nonlinear behavior of
various structural elements i.e., beams, plates, and shells, with various
boundary conditions, loaded statically and dynamically. Thousands of pa-
pers, books, and reports have been published on these subjects.
Purchased from American Institute of Aeronautics and Astronautics

xiv G.MAYMON

2) The second uses analytical and experimental models for various failure
criteria, allowable stress envelopes, accumulated damage effects, and crack
propagation processes. The introduction of fiber-reinforced composite ma-
terials enhanced these research efforts, because of the more complex failure
process of these materials.
3) The third uses algorithms developed for numerical computations of
structures. In the last 3Gyears tremendous research and development efforts
were invested in developing large finite element programs, which reduce
the need to apply approximate analytical solutions. Today, programs like
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

NASTRAN, ANSYS, and others are basic tools in the industry.


This direction is not explored in this book, although results that are
required for the understanding of the content are introduced. To obtain
in-depth information on this subject, the interested reader should explore
the existing literature.
Uncertainties in the loads has been treated by academic institutions and
industry during the last 20-30 years. Parameters of spectral density of
external excitations (such as acoustic noise, wind loadings, ocean waves, and
earthquake loads) were defined, using accumulated realistic data. Statistical
methods for analysis of structures under random excitations were intro-
duced into most of the qualification standards and into the design codes.
These methods were also introduced into the commercial finite element
programs, which are now used extensively in the industry. Chapters 1-7
deal with the resDonse of deterministic structures to random excitation.
Subjects covered include single-degree-of-freedom systems, multiple-de-
grees-of-freedom systems, continuous structures, and the treatment of non-
linear structures.
The response of an elastic structure to external dynamic loads plays an
important role in the total analysis of a designed structure. The loads acting
on an aerospace structure are both static and dynamic. Whereas the static
(or very slowly varying) loads create static deflections and stresses, the
structural response to dynamic loads are vibrations-time-varying displace-
ments and stresses. In most practical cases the loads are not deterministic
and have some statistical dispersion. A complete analysis of a practical
structure should involve bothstatic and dynamic behavior. In many cases,
as a result of the life history of the structure, both static and dynamic loads
act simultaneously, and this should be taken into account during the design
process. Resonance of the structure is sometimes responsible for high values
of response, even for small dynamic loads.
Aerospace structures are excited by aerodynamic loads, which are usually
random in nature. Flow around a structure creates pressure fluctuations,
which have a wide range of frequency and amplitude content. The same
phenomena are caused by acoustic noise created by rocket and jet outlet
flows. Rotating elements such as engines and rotors create excitations with
a better defined frequency content that are in many cases random in ampli-
tudes.
The dynamic response of a structure to any loads can be analyzed by
many ways such as energy methods, differential equations methods, and
integral equations solutions. The most common method for solving the
Purchased from American Institute of Aeronautics and Astronautics

PREFACE xv

response of practical structures to external excitations is the normal mode


approach, which is based on a superposition of normal modes that are
assumed to be characteristic of the structure. By its nature, the normal
modes superposition can be applied to linear structures, although nonlinear
cases can be treated by equivalent linear systems, as shown later in this
book. The normal mode method is effective and therefore popular, because
the normal modes and natural frequencies of the structure can be calculated
without any reference to the external excitation (thus, they characterize
the structure) and each mode can be treated separately as a single-degree-
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

of-freedom oscillator. The computation methods for the determination of


the normal modes and natural frequencies are well established, and all the
commercially available finite element computer codes include routines for
their computation for the practical cases where an analytic solution does
not exist.
Uncertainties in the structural parameters, or stochastic structures, have
been treated more extensively only during the last 10-20 years and are
described better in the introduction to Chapter 8. Chapters 8-13 deal with
various aspects of stochastic structures.
Problems of the response of a deterministic structure to random excita-
tions can now be solved by most of the commercially available finite element
codes. It could be asked: Why should I bother with mathematics, equations,
and approximate solutions when data can be keyed into the computer and
the program can provide the required answers? It is the author's opinion,
supported by more than 35 years of practical experience, that the computer
must be treated as a tool to speed up the task and not as a replacement
for the human brain. Knowledge of the basic assumptions, the capability
to quickly solve simplified models during the preliminary design stage (and
compare them to computer answers, like a process of calibration), the
ability to analyze results and draw smart conclusions, and the confidence
to decide on necessary changes to the design depend on the physical under-
standing of the phenomena involved. All these are the benefits of engineers
who do not let the computer think for them.
The understanding of the basic concepts is crucial to a correct method
of working. Therefore, in Chapter 1 the basic behavior of a single-degree-
of-freedom system is described. Basic behavior of a multiple-degrees-of-
freedom system is described in Chapter 2, and continuous structures are
detailed in Chapter 3. In Chapter 4, the reader is introduced to some
guidelines of random functions and the representation of random excita-
tions in specifications. Also included is a model for the excitation generated
by the boundary-layer flow over flat and lightly curved surfaces. The re-
sponse of a single-degree-of-freedom system, a multiple-degrees-of-free-
dom system, and a structure to random excitations is presented in Chapter 5.
In Chapter 6, behavior of a geometric nonlinear oscillator is described,
and in Chapter 7 the statistical linearization method is applied to demon-
strate the response of nonlinear single-degree-of-freedom systems, multi-
ple-degrees-of-freedom systems, and continuous structures to stationary
random excitations. It is shown that use of nonlinear analysis may often
result in a less conservative structure.
Purchased from American Institute of Aeronautics and Astronautics

xvi G. MAYMON

In Chapter 8, the basic concepts of the analysis of nondeterministic


structures are outlined and demonstrated. A possible analytical method
for the solution of static stochastic structures is described in Chapter 9.
Chapter 10 describes a Taylor series expansion method, which can be used
for the solution of more complex cases. Chapter 11 describes the modified
joint probability density function method, which enables the use of an
existing, commercially available finite element computer code to solve sto-
chastic structures problems. In Chapter 12 some aspects of dynamic behav-
ior of stochastic structures are described, and in Chapter 13, some aspects
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

of a stochastic crack propagation model are shown.


The assumptions and the most important practical conclusions are sum-
marized at the end of each chapter.
Mathematical proofs are not always included as these are often less
important to the practical user who, in most cases, is interested in the final
practical result. Nevertheless, the more interested reader is encouraged to
explore the cited references.
It is believed that design engineers who read this book and understand
the included examples will be able to extend the knowledge gained by this
reading to practical problems they encounter during daily engineering work.
Greater understanding of the physical meaning of the solutions will un-
doubtedly result in a much better designer.
Giora Maymon
October 1997
Purchased from American Institute of Aeronautics and Astronautics

Chapter 1

Deterministic Single-Degree-of-Freedom System


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. SDOF System Subjected to External Harmonic Excitation

A GOOD physical understanding of the behavior of vibrating structures


can be achieved by analyzing the behavior of an single-degree-of-
freedom (SDOF) oscillator. The SDOF system is covered by an extremely
large number of textbooks, e.g., Refs. 1-6, and is the basis for every course
in vibration analysis. It will be discussed briefly in the first chapters to
create a baseline for the analysis of random vibration.
The classical oscillator contains a point mass rn, i.e., all of the mass is
concentrated in a single point, which is connected to a rigid support through
two elements: a linear massless spring with stiffness k and a viscous damper
c (which creates a force proportional to the velocity) or a structural damper
h (which creates a force proportional to the displacement and in a 90-deg
phase lag behind it). The system can be excited either by a force f acting
on the mass or by a base movement x,. This type of system is described in
Fig. 1.1. Note that in Fig. 1.1 two elements connect the mass to the support.
This is only a schematic representation. The spring element represents the
stiffness of the structure, and the internal (structural) damping is repre-
sented by the viscous damping.
In the following evaluation, the classical viscous damping is assumed
(although treatment of structural damping is similar and can be found in
many references). The basic equation of motion is

The natural frequency of the undamped system is

The system is excited by a harmonic force of amplitude fo and frequency R

Assume a solution in the form


Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 1.1 SDOF system subjected to excitation force f ( t ) .

which yields the following result for xo:

where

This expression is complex, i.e., there is a phase difference between the


excitation force and the displacement. Using classical solution methods,
one can obtain the absolute value of xo,

where

where 5 is the damping coefficient given by

and is used extensively in engineering applications. The phase angle be-


tween the displacement and the force is
Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC SDOF SYSTEM 3

Note that in Eq. (1.6), folmwi = folk is the static deflection xStaticof the
SDOF system under a static force of magnitude& and, therefore, a dynamic
load factor (DLF) can be defined by
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The DLF expresses the dynamic amplification of a static deflection. A plot


of Eq. (1.9) is shown in Fig. 1.2. These are the well-known resonance curves.
Examination of Eq. (1.8) shows that, at very low-excitation frequencies
(SL/oo+ 0), x is in phase with the excitation. At resonance (fl/oo = 1) the
displacement is 90 deg ahead of the excitation, and in very high-excitation
frequencies (SL/oo 4 a ) there is a 180-deg lag between the displacement
and the force, or antiphase. In Fig. 1.3 the vectors f,x, f, and Y are shown
for R/oo -+ 0, SL/wo = 1, and fl/oo -+ a . Note that, in resonance, the
velocity vector is in phase with the excitation force. This result is used in
many experimental methods for the determination of resonance frequencies
and modes shapes, e.g., Refs. 7 and 8.

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8


Excitation to Resonance Frequencies Ratio
Fig. 1.2 Dynamic load factor.
Purchased from American Institute of Aeronautics and Astronautics

G.MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 1.3 Phase between force vector and displacement, velocity, and acceleration.

11. SDOF System Subjected to Base Excitation


Many engineering applications involve base excitation rather than force
excitation. Such are the cases of structures subjected to earthquakes, vehi-
cles moving on rough roads, and structural subsystems mounted on a main
structure. It is convenient to express the excitation by an input acceleration
xs(t) as shown in Fig. 1.4.
It can be shown that, when u = x(t) - xs(t), the relative displacement
between the mass and the support obeys the following differential equation:

mii + czi + ku = -mfs(t) (1.10)

In many cases, it is the relative displacement between the mass and the
support, and not the absolute displacement of the mass, that is responsible
for the stresses in a spring (or in a structure). Equation (1.10) describes a
system that is equivalent to that described in Eq. (1.1), with an equivalent
force equal to the mass multiplied by the base acceleration, in a direction
opposite to the base excitation. Thus, when xs = xsoei"', the term (-mfSo)
can replace fo in Eq. (1.6) whereas u, the relative displacement, replaces

Fig. 1.4 SDOF system subjected to base excitation.


Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC SDOF SYSTEM

x; thus,

whereas the expression for the phase angle remains unchanged.


When the support is moved by a harmonic displacement (rather than by
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

acceleration) of amplitude xsoand frequency a

the acceleration is obtained by double differentiation with respect to the


time t

and, therefore,

Equation (1.13) is described in Fig. 1.5.

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8


Excitation to Resonance Frequencies Ratio

Fig. 1.5 Amplification of base displacement.


Purchased from American Institute of Aeronautics and Astronautics

G.MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 1.6 Phase between excitation and relative displacement, velocity, and acceler-
ation.

The equivalent excitation force has a negative sign and, therefore, the
vector plots of x,, is,
$, u, ir, and ii are as described in Fig. 1.6.
For very low base excitation frequencies u + 0 (no relative displacement
occurs between the moving base and the mass and they move together).
For very high base excitation frequencies u tends to xso,resulting in a
relative displacement of -xSo,i.e., the mass does not move relative to the
external world and all of the relative movement between the base and the
mass is due to the base movement.

111. Response of an SDOF System to General Force


The most general method to find the response of an SDOF system to a
general force input is to use the response h(t) of such a system to a unit
impulse S(t). A unit impulse is defined as an infinitely large force acting
during an infinitely small period, so that the total impulse is one unit. The
solution for h(t) input is described in most of the available textbooks, and
the result for an SDOF system initially at rest, with damping coefficient
< 1 (which is usually the case in structural analysis) is
Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC SDOF SYSTEM 7

The response to a general force f (t) can be represented by a series of


repeated impulses of magnitude ~ ( T ) A Tapplied
, at t = T. In vibration
textbooks it is shown that the displacement of the SDOF system can be
written by either
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Equations (1.15) are known as the convolution or Duhamel's integral. As


h(t - T) is identically zero for t < T (the time instants preceding the
excitation of the system), Eqs. (1.15) can also be written as

where, for a system initially at rest, h ( r ) is given by Eq. (1.14).

Example 1.1
The use of Duhamel's integral (1.15b) is demonstrated for an SDOF
system with resonance frequency ooand a damping ratio l,subjected to a
step load, e.g., a load fo that is applied at t = 0 and remains constant. For
this case

1
h(r) =
mw, rne-@oTsin ( w o r n T)

Substituting these expressions into Eq. (1.15b) yields

Evaluation of this integral yields the following expression for the DLF:

This expression is described in Fig. 1.7.


Purchased from American Institute of Aeronautics and Astronautics

8 G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5


Time (sec)
Fig. 1.7 DLF for the example.

The integrals of Eqs. (1.15) and (1.16) cannot always be evaluated explic-
itly but numerical integration can be performed to determine the response
of an SDOF system to a general force. In Fig. 1.8 a flow chart that enables
simple programming of a procedure for this type of integration is shown.
When an analysis of random vibration of structures is performed, it is
common to d o many of the calculations in the frequency domain and not
in the time domain. Definition of external random inputs is usually given
in the former. Therefore, it is useful to demonstrate the evaluation of the
behavior of a SDOF system in the frequency domain and to demonstrate
the transformations between it and the time domain.
Equation (1.1) can be written as

where L2(dldt) is a second-order differential operator


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Purchased from American Institute of Aeronautics and Astronautics


Purchased from American Institute of Aeronautics and Astronautics

10 G. MAYMON

When f (t) is an harmonic excitation

f(t) =foei"'

A solution in the following form is sought:

x(t) = X(R)eint
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Substitution of Eqs. (1.18) and (1.19) into Eq. (1.17) yields

as dldt = iR.
Denoting

one obtains

where H(R) (called either the complex frequency response or the recep-
tance) is given by

which is identical to the definition given in Eq. (1.5).


The relation between the impulse response h(t) and complex frequency
response H(R) can be determined by the following procedure.
The general force f (t) is represented by a Fourier integral

f(t) = !+- F (R) ei"' dR (1.23)


-m

where

F(R) is a function equivalent to the Fourier coefficients in a Fourier series.


Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC SDOF SYSTEM 11

The response x ( t ) can also be represented as a Fourier integral


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Substitution of Eqs. (1.23) and (1.24) into Eq. (1.17) yields

and x ( t ) is, according to Eq. (1.24),

When f ( t ) is a result of a unit impulse applied at t = 0

From Eq. (1.26) one obtains

Thus, the impulse response h ( t ) multiplied by 2n is a Fourier transform of


H ( R ) . In some textbooks, Eq. (1.23) is defined slightly differently, so that
the complex transfer function is the Fourier transform of h ( t ) and not of
2nh(t).Nevertheless, Eq. (1.28)are more convenient for use in the response
of a system to random excitations.

IV. Stresses in an SDOF System


Formally, the term "stresses" is not appropriate for an SDOF system.
Nevertheless, it is clear from the behavior of an SDOF system that an
internal force exists in the spring, proportional to the change in its length.
The spring is the single elastic element in the SDOF system, as a point
mass is assumed. This force creates stresses in the massless spring. It should
be borne in mind that changes of dimensions of elastic elements, created
by internal forces in the loaded system, are the reason for the existence of
stresses in a system. This subject is evaluated when vibration of continuous
systems such as structural elements is discussed.
Purchased from American Institute of Aeronautics and Astronautics

12 G.MAYMON

V. Summary
In this chapter, the behavior of an SDOF system was described. The
assumptions that are made are as follows.
1) The SDOF system is linear.
2) The mass is concentrated at one point.
3) The spring is massless.
4) A viscous damping is demonstrated, although the treatment of struc-
tural damping is similar and can be found in many textbooks, e.g., Ref. 3.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

It was shown that an SDOF system subjected to an harmonic force


excitation vibrates in amplitudes that are different than those resulting
from the static deflection of such a system. The largest amplifications are
obtained when the excitation frequency is in the vicinity of the resonance
frequency. Where the excitation frequency is much higher than the reso-
nance frequency, attenuation is obtained. The response has a phase angle
with the excitation force, and its magnitude depends on the ratio between
the excitation and resonance frequencies.
When base excitation is applied, an equivalent excitation force can be
calculated, and the problem can be solved using the force response expres-
sions.
Duhamel's integral enables the computation of the response of an SDOF
system to a general excitation force.
Relations between the complex frequency response and the impulse
response function are also presented.
Purchased from American Institute of Aeronautics and Astronautics

Chapter 2

Deterministic Multiple-Degree-of-FreedomSystem
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Differential Equations and Normal Modes of an MDOF System

A MULTIPLE-DEGREE-OF-FREEDOM (MDOF) system contains


several masses, interconnected by springs and dampers and excited
by several external forces and/or base excitations. The degree of the MDOF
is determined by the number of masses n. Each mass mi (i = 1 , 2 , . . . , n)
moves with a displacement xi. The following differential equation can be
written in a matrix form:

where [m] is an (n X n) mass matrix that contains a combination of the


discrete masses of the system, [c] is an (n X n) damping matrix, [k] is an
(n X n) stiffness matrix, {f(t)} is an (n X 1) exciting forces vector, and
{x(t)) is an (n X 1) vector of the displacements of the discrete masses.
A system of two masses, three springs, and three viscous dampers excited
by two external forces is shown in Fig. 2.1. The two masses ml and m2 have
two DOFs xl (t) and x2(t) respectively. The springs, dampers, and excitation
forces are also marked in Fig. 2.1. When ml is moved by xl and m2 is moved
by x2, the forces acting on these masses are shown in Fig. 2.2.
Two equilibrium equations can be written

which can be rewritten as

Equation (2.2b) can be compared to Eq. (2.1).


Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I
, I
,
XI (t) x? (t)
Fig. 2.1 Two-DOF system.

Consider the free undamped vibration of the system

[mI@(t)}+ [ k l { x ( t ) }= 0 (2.3)
and assume a solution

{ x(t ) } = { x } eiot (2.4)


Substitution of Eq. (2.4) into Eq. (2.3) yields

( [ k ]- w2[m]){x}
eio' = {0} (2.5)
The nontrivial solution of Eq. (2.5) exists only when the determinant

from which a polynomial equation of the order n for n values of


w2 :w:, . . . ,w:, . . . , is obtained. Substituting any of these frequencies
back into Eq. (2.5) yields a corresponding set of relative values for {x}.

Fig. 2.2 Forces acting on the masses.


Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC MDOF SYSTEM 15

This is called the ith mode shape, and it is a column vector {+i}.
The n mode
shapes are described by a matrix [4] in which every column corresponds to
the ith mode shape of the frequency mi.The matrix [4]is in the form
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where the first index is the mode number and the second index indicates
the system coordinate. For example, 42,3denotes the nodal displacement
of the second mode in the third DOF (coordinate) of the structure.
Assume that in the example described in Fig. 2.1, ml = 2 kg, m2 = 1 kg,
kl = 0.5 X lo6 Nlm, k2 = 1 X lo6 Nlm, and k3 = 0.5 X lo6 Nlm. Using
Eqs. (2.2b) and (2.6) one obtains

from which of the following equation is obtained for 02:

whose solutions are of= 0.32461 X 106 ( r a d l ~ )and ~ o$ = 1.92539 X


lo6 ( r a d l ~ )Using
~. (Eq. 2.5) with x2 = 1 the following mode shapes matrix
is obtained:

When more than two DOFs exist, one of the many commercially available
eigenvalues and eigenvectors programs can be used to calculate the frequen-
cies and the mode shapes.

11. Generalized Masses, Dampings, Rigidities, and Forces


The resonance frequencies and the mode shapes are characteristic of the
system and not of the loading. They depend only on the masses and rigidities
of the system and, therefore, are attractive for use in structural dynamics
analyses. The mode shapes (sometimes called normal modes) possess an
important property known as orthogonality. This means that
Purchased from American Institute of Aeronautics and Astronautics

16 G.MAYMON

where the superscript Trepresents a transposed matrix and [MI is a diagonal


matrix with elements known as the generalized masses. In a similar way

and as a result of Eq. (2.5)

[ K ] = [o?M] (2.9b)
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where the elements of the diagonal matrix [ K ] are called generalized stiff-
nesses.
For a system with proportional damping (damping that is pr~portional
to the mass and/or the stiffness)

where the elements of the diagonal matrix [C] are called the generalized
dampings.
The assumption of proportional damping that yields a diagonal general-
ized damping matrix is not necessarily exact. Nevertheless, this assumption
greatly simplifies the calculations of structural response and, therefore, is
used extensively in practical engineering applications. When nodal damping
coefficients & are known (from experiments or from accumulated practical
knowledge) it can be shown that

It should be noted that the generalized quantities M , K , and C are not


unique. They depend on the values of [+I, which is a relative set of displace-
ments. Selection of [+] determines the generalized quantities. In some
cases, a normal mode is selected so that its maximum value is 1.This implies
certain generalized quantities. In other cases (usually in the large finite
element programs), [+I is selected so that all Mi = 1. It is not important
which normalization is made, as long as the process is consistent throughout
the whole solution. For the example solved earlier

is equivalent to another modal matrix


Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC MDOF SYSTEM 17

where both terms of the first column were divided by 1.17539. For the first
modal matrix,

and for the second


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

A very useful transformation, which is used extensively in all structural


dynamics analyses, is

i.e., the displacement vector {x(t)}is expressed in terms of the normal modes
[4] and a generalized coordinates vector { ~ ( t ) )As. this transformation is
used extensively in future chapters, it will be demonstrated in more detail
for the two-DOF system.
The modal matrix [4] of this system is

Equation (2.12), therefore, can be written as

or explicitly

In Eq. (2.14b) the displacements x, are expressed as a linear combination


of the generalized coordinates qi, weighted by the modal shapes.
Substituting Eq. (2.14b) into Eq. (2.1) and premultiplying each term with
[$] yields
Purchased from American Institute of Aeronautics and Astronautics

18 G.MAYMON

The quantity on the right-hand side of Eq. (2.15b) is called the generalized
forces matrix [ F]. For the preceding example
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Therefore, the generalized forces represent the work done by the external
forces when the masses of the system move a modal displacement. When
the normal modes are known, the generalized forces { F ) can be easily
calculated for a given set of external forces {f}.

111. Uncoupled Differential Equations


Writing Eq. (2.15b) explicitly yields

Mniin + Cniln + K n ~ n= Fn
Using Eqs. (2.9b) and (2.11), Eq. (2.17) can be written as

M1iil-t 2llWMli71 + 4 M l r l l = Fl

Equations (2.18) represent a set of uncoupled differential equations for


the generalized coordinates vi. Each of these equations can be solved
separately for qi, using all of the techniques and procedures of an SDOF
system, which are well known and documented. Then the displacements
{ x ) can be calculated using Eq. (2.12). This is an important feature of the
normal modes. It allows the engineer to solve an MDOF system using
solutions of an SDOF one, once the normal modes, resonance frequencies,
generalized masses, and forces are calculated, and the modal damping
Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC MDOF SYSTEM 19

coefficients are known or assumed. It is shown in the next chapter that this
approach can also be used for continuous structures.

IV. Summary
Using equilibrium equations, the differential equations of an MDOF
system can be written. Resonance frequencies and mode shapes can then
be calculated.
Using the orthogonality of the normal modes, generalized masses and
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

other generalized quantities can be calculated using Eqs. (2.8), (2.9b),


and (2.11).
When the displacements are expressed as a linear combination of the
normal modes and general coordinates, the transformation yields a set of
uncoupled linear differential equations, and each of these equations can
be solved for the generalized coordinate 7,using expressions and procedures
evaluated for an SDOF system.
When an MDOF system is subjected to base excitations, the equivalent
excitation force can be calculated, using the expressions given in the next
chapter. Once these forces are known, the procedures of an SDOF system
can be used for the generalized coordinates, and the system displacements
can then be expressed using Eq. (2.12).
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

Chapter 3

Deterministic Continuous System


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Introduction

I N A continuous system, there are an infinitely large number of infinitely


small masses. The stiffness is contributed bv internal forces interacting"
between these infinitely small masses, and the damping is created by internal
dissipation of energy throughout the system. Therefore, this type of a
system has an infinite number of DOFs. A practical structure built of
interconnected beams, plates, and shells is atypical continuous system.
Usually there are no discrete masses, although in some engineering applica-
tions the distributed mass is lumped into discrete masses.
It has been shown in many textbooks, e.g., Refs. 1-6, that a continuous
system has an infinite number of resonance frequencies and normal modes.
For many simple structures such as beams with various boundary conditions,
rectangular and circular plates, and simple shells, analytical expressions
can be evaluated for these frequencies and mode shapes. On the other hand,
it is seldom possible to find closed-form expressions for these properties for
a practical structure in which different kinds of structural elements and
realistic boundary conditions exist.
For these reasons resonance frequencies and normal modes of a practical
engineering structures are calculated numerically. The large commercial
finite element computer codes (such as NASTRAN and ANSYS) are capa-
ble of computing these structural properties. As the basic concept of finite
elements is the discretization of a continuous system into a large number
of interconnected elements, these computer codes actually treat an MDOF
system. Instead of an infinite number of resonance frequencies and normal
modes, a finite (though very large) number of resonances are computed.
Experience has shown that a practical continuous system can be treated as
a large MDOF system with sufficient engineering accuracy.
Suppose that the mode shapes {4i} (i = 1 , 2 , . . . , m) and the resonance
frequencies oi(i = 1,2, . . . , m) were calculated. Then, the deflection w(x,
t ) of the structure can be exvressed in terms of the normal modes and the
generalized coordinates qi(t),

It should be noted that x in Eq. (3.la) does not necessarily represent a


one-dimensional structure, but a location in the structure as follows: one-
Purchased from American Institute of Aeronautics and Astronautics

22 G.MAYMON

dimensional coordinate x for a one-dimensional structure (such as a beam


or a frame), two-dimensional coordinates x, y or r, 8 for a two-dimensional
structure (such as flat and curved plates and shells of revolution), and three-
dimensional coordinate x, y, z or r, 8, 9 for a three-dimensional structure
(such as solids).
If the modes were solved numerically for N DOFs, Eq. (3.la) takes
the form
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

11. Differential Equations of a Continuous System


The evaluation of the differential equations of motion of a continuous
system are done using the Lagrange equations. For a system in which the
damping forces are derived from a dissipation function D, the rth Lagrange
equation is given by

where T is the kinetic energy of the system, U is the potential energy, D


is the dissipation function, and N is the work done by the external force
in the generalized coordinate.
For clarity, the development of the equations of motion is done for a
uniform beam, without any loss of generality. For such a beam, m is the
mass per unit length, EZ is the flexural stiffness, c is the damping per unit
length, and w(x, t ) is the displacement normal to the structure along the
x axis. The kinetic and potential energies are

and the dissipation function is

where proportional damping is assumed.


From Eq. (3.lb)
Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC CONTINUOUS SYSTEM

From Eq. (3.2)

Changing the order of integration and summation yields


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Being normal modes, 4iand +, are orthogonal; thus,


J rn+,(x)+i(x) dx = 0 for i # r

J rn+,(x)+i(x) d x = Mr for i =r (3.10)

where Mr is the rth generalized mass. Equation (3.8) takes the form

If the notation described in Chapter 1 is applied for the damping, using a


modal damping coefficient, then

c = 2[,wrm (3.12)

Substituting Eq. (3.12) into Eq. (3.5) yields

D = 25,w, '/z $ mw2 dx = 25,w,T (3.13)

and, therefore,

Using a similar procedure for the stiffness yields

To find an expression for the work done by the external forces, assume
that a distributed force per unit length p(x, t ) is applied to the beam. The
virtual work done by this force in a virtual displacement Sw(x, t ) is
Purchased from American Institute of Aeronautics and Astronautics

24 G.MAYMON

From Eq. (3.1a)

and, therefore,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This should be equal to the work done by N, on a virtual displacement Sqi

Comparing Eqs. (3.18) and (3.19) yields

For n discrete forces fn(t) rather than a distributed force, acting at points
x = x n ,it is possible to show by the same procedure that

Thus, Ni (or Nr) is a generalized force, defined exactly as for the MDOF
system (Chapter 2) is the work done by the external forces when a modal
displacement +i is applied to the system.
Substituting Eqs. (3.11), (3.14), and (3.15) into Eq. (3.2) yields a set of
N uncoupled differential equations for the generalized coordinates

When an infinite number of DOFs is assumed, the set (3.21) contains an


infinite number of equations.
Equations (3.21) are identical to Eq. (2.18). Therefore, the procedure
for computation of a deterministic response of a continuous system is
as follows:
1) Decide how many DOFs are to replace the infinite number of a given
structure. This decision must be based on the nature of the structure and
should take into account the engineer's experience in designing similar
structures.
2) Calculate the resonance frequencies and mode shapes. This can be
done analytically for simple structural elements or numerically using a finite
element computer program.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Purchased from American Institute of Aeronautics and Astronautics


Purchased from American Institute of Aeronautics and Astronautics

26 G. MAYMON

For one-dimensional structures, i.e., beams, with mass per unit length
m(x) and load of force per unit length q(x)

where the integral is taken along the whole length of the structure.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

For two-dimensional structures, i.e., plates, with mass per unit area
m(x, y) and load of force per unit area p(x, y)

where the integral is taken over the whole surface of the structure.
For three-dimensional structures, i.e., solids, with mass per unit volume
p(x) and load of force per unit volume s(x)

where the integral is taken over the whole volume of the structure.

111. Base Excitation


Many practical engineering problems are characterized by an input that
is a base excitation rather than an external force. Examples of such cases
are a vehicle traveling on a rough road, a structure excited by ground
movement (earthquake), or a substructure mounted on a vibrating main
structure. In Chapter 1 it was shown for an SDOF system that the equation
of motion contains an equivalent excitation force, which is related to the
base motion. The formulation of these equivalent forces for a continuous
structure is best demonstrated on a one-dimensional structure, with two
spatial coordinates, such as a frame.
Assume that the frame shown in Fig. 3.2 vibrates in a mode also described
in the figure. The spatial coordinate s runs along the frame. The rth normal
mode 4,(s) runs along the structure. This mode has a component 4,,(s)
in the x direction and +,,(s) in the y direction. The displacement of each
point on the structure is then a combination of the elastic displacement
Z(S,t )
Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC CONTINUOUS SYSTEM


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 3.2 Frame subjected to base excitation.

and the rigid body displacements wo, and w ~ ,thus,


~;

The kinetic energy of the system is

where m ( s ) is the mass per unit length of the frame and m(s) ds is the
total mass of a small element of length ds.
From Eq. (3.26)
Purchased from American Institute of Aeronautics and Astronautics

28 G.MAYMON

and, therefore,

ailr

Because of the orthogonality of the normal modes and the relation


+f( s ) = + f T + +f,,,the following expression is obtained, after some algebra:
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

and, therefore,

This expression can be identified as part of the Lagrange equations of


motion [see Eq. (3.2)] originating from the kinetic energy of the system.
As the potential and damping energies depend entirely on the elastic defor-
mation of the structure, the only addition to the Lagrange equations of
motion are the first and second terms on the right-hand side of Eq. (3.31).
Thus, the equation of motion of the rth mode is

The equivalent generalized force is the base acceleration multiplied by the


mass weighted by the normal modes in a direction opposite to the base
excitation. When, for example, base excitation in the y direction exists, the
integral J m ( ~ ) + ~ , ' (ds
s ) can be calculated and the equivalent excitation
generalized force is (-wo,, $ m ( ~ ) + ~ , ~ d s()s. )
The external base movement can be given as displacement excitation,
i.e., the case of a vehicle traveling on a rough road, or as acceleration, i.e.,
earthquake. If harmonic displacement is imposed, say in they direction, then

w o , = A, sin flt (3.33a)

and the equivalent generalized force is a function of f12, because

w o , = -Ayf12 sin a t (3.33b)


Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC CONTINUOUS SYSTEM 29

On the other hand, if a harmonic acceleration is imposed, then

w ~= ,ay~sin flt (3.34)

and the right-hand side of Eq. (3.32) does not include f12.
Once again it is demonstrated that a vibration problem of a practical
structure of an infinitely large number of DOFs can be treated as a collection
of SDOF systems.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

IV. Stress Response


Stresses in elastic structures are uniquely defined by the deflection, using
the material constitutive relations and the compatibility equations. The
means by which the deflection was obtained, statically or dynamically, is
irrelevant. The effect of the dynamic load factor or the amplification factor,
demonstrated in Chapter 1 for an SDOF system, is introduced into the
system when the deflections are calculated.
For a specific type of structure, the following relation is always present:

where Liis a differential operator in the spatial coordinate, K is a constant


that depends on the elastic constants and the geometry of the structure,
and i is an index that indicates which stress is being calculated. For instance,
it can be arbitrarily defined that ui= axis axial stress in the x direction,
ai= uy is bending stress in the y direction, ui= r,, is shear stress in the
structure, ui= uxy is shear stress between two plies in a laminated structure,
and ai= us,is stress at the edge of a hole in a plate.
Two examples for K and Liin Eq. (3.35) are as follows.
1) The bending stress in the tensed side of a deflected beam is

2) The bending stress in the x direction on the tensed side of an isotropic


plate is

Because it was defined in Eq. (3.la) that


Purchased from American Institute of Aeronautics and Astronautics

30 G. MAYMON

then

and according to Eq. (3.35)


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

which can be rewritten as

where

Equation (3.39) resembles Eq. (3.36). Therefore, the function ?P\Ir,(')(x)was


designated by the author in Ref. 9 as stress mode. By the use of the stress
modes, an evaluation of the stress response of a structure can be repeated
in the same way it was evaluated for the displacement response.
Because the normal modes $Ax) are characteristics of the structure, it
can be seen from Eq. (3.40) that the stress modes are also characteristic
to the structure and not to the loading condition. Once stress modes for a
given structure are computed, a complete mapping of stress distribution in
the randomly vibrating structure can be performed. To obtain the stress
modes, one naturally tends to use Eq. (3.40), i.e., to apply the operator L
on the mode shapes. There are several disadvantages in doing so.
1) The operator L is not always known in a closed-form expression. The
examples given earlier were related to very simple cases.
2) The mode shapes of the structure are not always available in a closed-
form expression. In many cases, mode shapes are approximated by assumed
functions that satisfy the boundary conditions but are still only approxima-
tions. Whereas the use of assumed modes can yield accurate results for the
frequencies and deflections of a structure, their differentiation (usually
double differentiation is required to obtain stress) can introduce signifi-
cantly large errors in the stress results.
3) If mode shapes are obtained by a numerical solution such as a finite
element computer code (a procedure that is a routine in practical engi-
neering solutions) numerical differentiation of these modes will introduce
large errors in the stress results.
For these reasons, a more practical method must be found for the determi-
nation of the stress modes. Examination of Eqs. (3.39) and (3.40) shows
clearly that the stress mode is the stress distribution in the structure when
Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC CONTINUOUS SYSTEM 31

the latter is subjected to a deflection equal to the mode shape, with the
generalized coordinate normalized to one. This observation leads to another
practical definition of the stress mode: The stress mode is the stress distribu-
tion in the structure, when the latter is deformed statically into a deflection
equal to the mode shape +,(x).
The application of this definition is very simple. Most of the commercially
available finite element codes can calculate stresses for an initially imposed
static deflection. Therefore, the user introduces an imposed deflection to
the structure with a distribution equal to the normal mode, and the required
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

stress mode of kind i (which may vary according to the specific problem)
is calculated statically.
The use of stress modes calculated statistically brings the designer to a
better understanding of the behavior of the structure. Past experience has
shown that a design engineer usually has a good intuitive feel for static
loads and is able to identify points of weakness in a design by looking at
a static analysis. This feel is less reliable where dynamic stresses are con-
cerned. In these cases, the meaning of the normal modes may be well
understood, but it is difficult to visualize a physical interpretation of a
weighted combination of these modes. It is even more difficult to feel
random vibration, mainly because the excitation includes many compo-
nents, each with a different frequency. It is interesting to note that the
structural locations that experience the largest amplitudes are not necessar-
ily those in which maximum stresses exist. The classical example is the
cantilever beam, in which the largest amplitude is obtained at the free tip,
while the largest stress is located at the clamped edge, where the amplitude
is zero. By inspecting the stress modes, a better understanding of the dy-
namic stress distribution in the structure is obtained.
When calculating the stress modes, care must be taken to determine the
dimensions of the quantities involved in the process. In Eq. (3.36) the
normal mode was assumed to be dimensionless, and the dimension of the
deflection w was introduced through the general coordinate 17(t). Examina-
tion of Eq. (3.39) reveals that to obtain stresses in the correct dimension,
the stress mode must have a dimension of (stressllength). This adjustment
is required when using the procedure described for the determination of
the stress modes and dimensional deflection must be introduced into the
numerical algorithm. Thus, the introduced deflection is not +,(x), but rather
Ao+,(x), where A. is a unit displacement of magnitude 1 and dimension
of length. Therefore, the result of the numerical calculation is not
*\Irji)(x) but Ao!P\i)(x).Practically, the numerical values are the same, but
the dimensions are divided by a dimension of length. The preceding argu-
mentation is important when mixed dimensions of length are used, e.g.,
length in centimeters and meters, or inches and feet. It is a good engineering
practice to perform the entire analysis throughout the solution of a given
structure with one set of dimensions in length, mass or force, and time.
When this is done, no problems arise during the interpretation of the results.
Two simple examples of a stress mode and its dimensions are demon-
strated.
Purchased from American Institute of Aeronautics and Astronautics

32 G. MAYMON

Example 3.1
A simply supported beam of height h (cm) and length L (cm) is loaded
by a uniformly distributed force. Young's modulus of the material is E (kg/
cm2) and the cross section of the beam has a moment of inertia Z (cm4).
The first stress mode of the longitudinal bending stress on the tensed side
of the beam is required.
The first normal bending mode of the beam is
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

nx
~ $ ~ (=
x )sin -
L

and the relation between the bending moment and the deflection is

The relation between the bending moment and the bending stress is

Assume that the beam is bent into the first mode

nx
w( x ) = A. sin -
L

where A, = 1 cm. Then

*\bend) n2Eh sin ?IX [(kg/cm2)/(cm)]


=-
2L2 L

Example 3.2
A simply supported rectangular plate of length a (cm), width b (cm),
and thickness h (cm) is loaded by a uniform pressure q. Young's modulus
of the material is E(kglcm2) and Poisson's coefficient is v (dimensionless).
The first stress mode of bending stress in the x direction is required.
Purchased from American Institute of Aeronautics and Astronautics

DETERMINISTIC CONTINUOUS SYSTEM 33

The first normal bending mode of a simply supported rectangular plate is

(x, y) 7ix
= sin - . sin -
?Ty
a b

The closed-form relation between the bending stress in the x direction and
the deflection is
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where D = Eh3/12(1 - v2) was substituted.


Assume that the plate is bent into a deflection

w = A0 . +1(x,y)
where A. = 1 cm, then

therefore,

9F e n d ) =
r2Eh
2(1 - v2) a
[++ $1 sin
7ix
sin
7iy
[(kg/cm2)/(cm)]

When these calculations are carried out with a finite element code, and
not by a closed-form expression, the stress modes are defined by a vector
of values of the stress modes calculated at the nodal points of the finite
element model. A simple example of this procedure is demonstrated by a
numerical analysis of a beam with linearly varied height, shown in Fig. 3.3.
The beam is made of a material with E = 2.1 X lo6 kg/cm2 and density of
p = 7.959 x kg s2/cm4.
The first mode shape is calculated using the ANSYS finite element pro-
gram. The beam is divided into 10 two-dimensional beam elements and

Fig. 3.3 Beam with linearly varied height.


Purchased from American Institute of Aeronautics and Astronautics

34 G. MAYMON

calculated with the modal analysis module of the ANSYS. The first fre-
quency obtained is ol = 168.74 radls, and the mode shape for 11 nodal
points is
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Next, a static case is solved by the ANSYS program, where a deflection of


the described mode shape (in centimeters) is forced. The stress mode of
the bending stress on the lower outer surface of the beam is

and the dimensions of this stress mode are in kilograms per square centime-
ters per centimeter.

V. Summary
MDOF and continuous systems for which resonance frequencies and
mode shapes were calculated can be represented by a set of uncoupled
differential equations that include generalized masses and generalized
forces. These equations are then solved as a set of SDOF systems.
The treatment of base excitation of a continuous system is similar to that
of the MDOF system, where an equivalent generalized force replaces the
force term in the regular vibration differential equation.
The concept of stress modes is presented. Use of these modes enables
an easy computation of the stress response of both MDOF and continu-
ous systems.
It is again emphasized that good knowledge and experience in solving an
SDOF system enables an easy solution of MDOF and continuous systems.
Purchased from American Institute of Aeronautics and Astronautics

Chapter 4

Random Functions and Excitation


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Basic Concepts of ,Random Functions

M ANY practical structures in aeronautical, mechanical, and civil engi-


neering are excited by forces, pressures, and base movements that are
random in time. In many cases the spatial distribution of these excitations is
also random. The time history of these excitations are not known determin-
istically. Nevertheless many global parameters of random excitation can
be determined, including expected values, mean values, standard deviations,
probability density functions of the magnitude of the excitations and their
peaks, and average zero crossing rates, as well as the relations between an
expected value at a certain time and the value at another time.
It is not the purpose of this chapter either to repeat the basic theories
of random variables, fields, and processes, or to prove basic mathematical
relationships. These are well documented in numerous books, e.g., Refs.
10 and 11, papers, and reports. Only basic definitions from the theory of
random functions and some practical aspects of these theories that are
important to the user are repeated here. Reference 11 and its cited refer-
ences is recommended for the interested reader.
Assume X(t) is a random function in time with random values at different
times. The probability distribution function of X(t), Fx(x; t) [sometimes
called the cumulative distribution function (CDF)] is defined as

i.e., this is the probability that X(t) is smaller or equal to a given value x.
The probability density function (PDF) is defined as

Note that in Eqs. (4.1) and (4.2) the superscript of the random function is
written with upper case letter X, and a specific realization of the function
is written in lower case x.
Further details about CDF and PDF, as well as specific examples of these
functions, are described in any textbook on the theory of probability. In
Fig. 4.1, an example of CDF and PDF is demonstrated.
The mean value of the function X(t), m,(t), is defined as the expected
Purchased from American Institute of Aeronautics and Astronautics

36 G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

-1.00 -0.50 0.00 0.50 1.00 1.50 2.00 2.50 3.00


X
Fig. 4.1 Example of CDF and PDF.

value of X ( t ) , and is the first statistical moment of X ( t ) , which has a PDF


fx(x; t )

Thus, m x ( t )is the location of the center of the area or center of gravity of
the function described in Fig. 4.1, along the x axis.
The second moment of X ( t ) is

This second statistical moment is called the mean square value of the
function X ( t ) and is the moment of inertia of the area bounded by fx(x;
t ) around the axis x = 0. The second statistical moment of X ( t ) around
the mean value m x ( t )is

This quantity is called the variance of X ( t ) and is the moment of inertia


of the area bounded by fx(x; t ) around the axis x = mx(t).The variance
is a measure of the spread of X ( t ) around the mean value. The root mean
square of the variance ax= v ' v a r ( ~ ( t )is) known as the standard deviation.
Higher-order statistical moments can also be defined, and these defini-
tions are well documented in the literature.
Purchased from American Institute of Aeronautics and Astronautics

RANDOM FUNCTIONS AND EXCITATION 37

It can be easily shown that

Thus, when the mean value of X ( t ) is zero, its variance is equal to its mean
square value.
The autocorrelation function of X ( t ) is defined as the expected value of
the product of X at times tl and t2
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where E [ - ]is defined as the expected value of [ 1.


A random function X ( t ) is said to be stationary when its value in t2 =
tl + r depends only on the time difference 7 = t2 - t, between the two
time instants tl and t2. Then the autocorrelation function is

Note from Eq. (4.1) that when r = 0

Thus, the value of the autocorrelation function for zero time lag is the
mean square of the function.
When X ( t ) and Y ( t ) are two random functions, the cross-correlation
between X and Y is similarly defined as

Two stationary functions are called jointly stationary when their cross-
correlation function depends only on the time difference 7; thus,

The concept of stationary functions is important when continuous vibrations


of systems and structures are considered, as in this volume.
The spectral density of a stationary random function Sx(w) is defined as
the Fourier transform of its autocorrelation function R x ( r )

These relations, called the Wiener-Khintchine formulas, state that R x ( r )


and Sx(w) are interrelated throughout a Fourier transformation. Some
Purchased from American Institute of Aeronautics and Astronautics

38 G. MAYMON

textbooks present a slightly different formulation of these relations, which


may differ by a factor of 27.r. The formulation in Eq. (4.12) is convenient
because it can be easily shown that

thus, the mean square of a quantity X i s the integral of its spectral density
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

over the frequency range. It can be seen that the spectral density is a
function of the frequency. In effect, for a given frequency band, the spectral
density function is the contribution of the quantity X to the total mean
square of that quantity.
In a similar way, a cross-spectral density function can be defined as the
Fourier transform of the cross-correlation function.

Random excitations (forces, pressures, base movements) and random re-


sponses may be characterized by their autocorrelation and cross-correlation
functions or by their spectral and cross-spectral density functions.
The autocorrelation of an excitation ql is a function that describes the
expected value of the product of a load at point xl at time tl, ql(xl, tl),
and the load at point x2 at time t2, ql(x2, t2). It is defined as

When two excitations ql and q2 exist, the cross correlation describes the
expected value of the product of one function q, at point xl at time t,,
ql(xl, tl), and the load q2 at point x2 at time t2, q2(x2,t2). It is described as

Thus, when the given excitation is one force acting in one location xl it
can have an autocorrelation function, but no cross correlation. On the other
hand, when two or more forces are acting on a structure, there may be
cross correlation between them. An external pressure acting on a structure
may be thought of as an infinite number of forces, and cross-correlation
functions between pair of points may be found. There are many cases in
which there is a correlation between the loads on two or more points on
the structure. This correlation may also be time dependent. For instance,
a turbulent flow may cause a pressure fluctuation at point XI, which will
influence the pressure at another point x2. Also, as the turbulence is swept
Purchased from American Institute of Aeronautics and Astronautics

RANDOM FUNCTIONS AND EXCITATION 39

downstream, there will be a correlation between the pressure at point x2


at time t2 and the pressure that existed at point xl at an earlier time t,.
As a consequence of the earlier argumentations, a cross-spectral density
matrix can be constructed for a system that is excited by n forces, each
force having a spectral density function Sii(w), i = 1, 2, . . . , n, and each
pair of forces having a cross-spectral density function Sij(o), i = 1.2, . . . , n,
j = l , 2 , . . . , n,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where diagonal terms are spectral density functions and off-diagonal terms
are cross-spectral density functions.
When a system is excited by random forces, it responds with movements
at all of the points. It is quite reasonable to conclude that the response to
a random excitation is also random. Even when only one force is acting at
one location, response is created at all of the points. Therefore, responses
always have autocorrelation and cross-correlation functions, and spectral
density and cross-spectral density functions. A cross-spectral density matrix
similar to Eq. (4.16), therefore, can be constructed for the responses of n
points on the structure.

11. Practical Characterization of Random Excitation


In most practical cases, the external spectral densities of random excita-
tions are the parameters that are given to the designer as an input. An
example of this type of input are the military standards (such as Ref. 12),
which provide means to generate a spectral density input, either of sound
pressure levels, or of an acceleration, as function of the frequency. The
definition of the spectral density inputs to the structure or its components
is part of the data of the loads definition document prepared at the beginning
of any project. In general cases, standards and specifications are used. In
some other cases, the experience gained by a designer in previous designs,
together with experimental data obtained during development of earlier
projects, forms the basis for the definition of the input for the new design.
The dimensions of the spectral density function are ( q ~ a n t i t y ) ~ /
frequency. Thus, a spectral density of force will have the dimensions of
(ne~tons)~/(radian/second), or (kilogram)2/(radian/second), or (pound)2/
(radianlsecond). The spectral density of pressure will be (Pas~al)~/(radianI
second), or (atmo~phere)~I(radian/second), or (pounds per square inch)2/
(radianlsecond), and the spectral density of displacement is ent ti meter)^/
(radianlsecond), or (in~hes)~/(radian/second). In all these alternatives, the
dimension of the frequency is given in radianlsecond. It should be noted
Purchased from American Institute of Aeronautics and Astronautics

G.MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

fc oc=21rfc

Fig. 4.2 Wideband excitation, example.

that the basic dimension of an angle is a radian, and not a cycle, which is
27r rads. Therefore, it is important to avoid the confusion that sometimes
arises between radianlsecond and cyclelsecond (hertz). In practical cases,
the spectral density defined by specifications and standards is given by
(q~antity)~/hertz, or (quantity)2/(cycle/second). It should be noted that
irrespective of the presentation of the spectral density, the mean square of
the quantity, which is the integral of the spectral density with respect to
the frequency, must be identical in the two possible presentations, and must
have the dimension of (quantity)2.
Let Sof be the value of a wideband excitation (Fig. 4.2) of a force F, a
constant in (kilogram)2/(cycle/second) over a frequency range between
f = 0 cps and f = f, cps. The mean square of the excitation is the area
bounded by the spectral density function. For this case it is

In the same way, if So, is given in (kil~gram)~I(radian/second) over a


frequency range 0 5 o 5 o,, the mean square of the excitation force is

E{F2) = So, I*[


radls
* o,[radls] = So,wc[kg2]

and as a result

but wc = 2nfc and, therefore,


Purchased from American Institute of Aeronautics and Astronautics

RANDOM FUNCTIONS AND EXCITATION 41

This result can be obtained by substituting the relation cycle = 2a rads in


the dimension. For example, assume a constant spectral density of value 5
kg2/cpsbetween the frequencies 0 and 100 cps. This will be transformed into

5 kg2
- 5 kg2 5 -kg2
=-
cps 2a(rad/s) 2a (radls)

between the angular frequencies 0 and 100 . 2a radls.


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

In general, when the spectral density of the excitation is dictated by


specifications or standards the data is given in (quantity)'/hertz. To avoid
confusion when calculating responses, it is recommended that frequencies
are always transformed into radianlsecond, and the integrals in the fre-
quency domain are performed on o (radls), and not on f (cps or Hz).
Sometimes the excitation is a random acoustic pressure, which is usually
defined in the specifications by the sound pressure level (SPL) over a given
range of frequencies. In Appendix B, the conversion of acoustic SPLs
levels (usually given as acoustic decibel) into a spectral density of pressure
excitation is described.

111. Important Excitation Functions


Several power spectral density (PSD) functions are frequently used in
the analysis of the response of structures to random excitation. 1) A white
noise excitation is one that has a constant value Sx(o) = SO= const over
the whole frequency range, - w < w < + m. 2) A one-sided white noise is
one that has a constant value S,(w) = So = const over the whole positive
frequency range, 0 5 o < + m. 3) A band-limited white noise is one which
has a constant value Sx(o) = So = const over a limited range of frequencies,
-oc 5 w 5 w, and zero elsewhere. 4) A one-sided band-limited white noise
is one that has a constant value S,(w) = SO= const over a limited positive
range of frequencies, 0 5 o 5 o, and zero elsewhere. These four excitations
are shown in Fig. 4.3.
From these four constant value PSD functions, the one-sided band-
limited white noise is the most often used, because of its practical features.
It is important to note that when the excitation is identically zero over
part of the frequency range, the mean square value can be obtained by
integration only over the relevant part of the frequency axis. Thus, for a
band-limited white noise,

and for a one-sided band-limited white noise


Purchased from American Institute of Aeronautics and Astronautics

G.MAYMON

one-sided
band-limited band-limited
white noise white noise
I
-- -1- ----. -- - ----
I
I I
I I
I 1
I I
I I
I I
I I
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I I

I / -
I white noise one-sided I
1

white noise j
I

1
-Wc +Oc Frequency
0
Fig. 4.3 Four constant value PSD functions.

In C h a ~ t e 5r it is shown that calculation of the mean value of the remonse


of a linear system to random excitation involves integration of the input
PSD multiplied by expressions of the complex transfer functions and their
conjugates over the whole relevant frequency range. Some analytical results
for cases where the input PSD is constant are shown in Appendix A, for
the whole frequency range - a~ < w < + a ~ A . very important results shown
in the literature, e.g., Refs. 3 and 11, is that for a system with a small
damping ratio (say,-less than 0.1), which is a frequent case in structural
analysis, the result for a band-limited integration can be well approximated
by the unlimited band integrations shown in Appendix A, provided that
the cutoff frequency o, is sufficiently higher (at least 30%) than the reso-
nance frequency.
Another important excitation PSD function is obtained when a one-sided
white noise is filtered through an SDOF system that has a resonance w,
and damping 5,. This filtering results with the following PSD function:

Note that w, and 5, can be used as parameters that may be selected to


approximate many given excitation spectral density functions. Analytical
results for integration of Eq. (4.22) together with the relevant transfer
function are also included in Appendix A, based on the results of Ref. 13.
An example of the shape of Zo is shown in Fig. 4.4.
Purchased from American Institute of Aeronautics and Astronautics

RANDOM FUNCTIONS AND EXCITATION 43


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

0 2 4 6 8 10 12 14 16 18 20
Frequency (radlsec)

Fig. 4.4 Filtered one-sided white noise.

IV. Boundary-Layer Excitation Model


In flight vehicles, the origin of a major kind of excitation is the turbulent
flow that exists around the structure. The flow causes pressure fluctuations
on different surfaces of the structure. These pressure fluctuations, which
are random in nature, cause a vibration response of the flat and curved
panels. The formation of pressure fluctuations is characterized by a correla-
tion between the pressure at a given location and the pressure at other
locations. The pressure that is built up in one location is not discrete, and
other locations are influenced. Also, the turbulence that is formed in one
location is swept downstream to form a pressure fluctuation at another
point. It is reasonable to assume that the pressure in the latter is influenced
by the first turbulence and that this effect depends on the downstream ve-
locity.
To solve the problem of response to a correlated turbulent boundary
layer, the correlation functions or the cross spectral densities of the pressure
fluctuations are required. It should be remembered that there is a Fourier
transform relationship [the Wiener-Khintchine equations (4.12)] between
these two functions.
To determine the correlations and spectral densities of a turbulent bound-
ary layer, experimental results taken from wind-tunnel measurements are
Purchased from American Institute of Aeronautics and Astronautics

44 G.MAYMON

used. In the past, experiments for this purpose were perf~rrned,'~.'~ and
the proper functions, which contain a number of experimental constants,
were fitted to the measurements. The model of Crocker14 is used in this
chapter, together with measurements performed by Mae~trello.'~
Some basic concepts of turbulent boundary layer are repeated here for
completeness. Further details can be found in publications on aerodynamics,
e.g., Ref. 16.
A boundary layer is defined as the layer of the flow in which changes in
the flow velocity exist. These changes vary between zero velocity at the
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

surface and uo, the external flow speed, as shown in Fig. 4.5.
The thickness of the boundary layer is denoted as 8. The flow velocity
is uo,

where M is the flow Mach number and c is the speed of sound in the
unperturbed flow.

u
Fig. 4.5 Velocity in the boundary layer.
Purchased from American Institute of Aeronautics and Astronautics

RANDOM FUNCTIONS AND EXCITATION 45

A displacement thickness 6* of the boundary layer is defined by

The displacement thickness is a function of the shape of u ( y )in the bound-


ary layer. Experiments show that
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

In different models, different values of n are determined. In one of the


models n = 7 is used so that

In the range 0 5 y 5 6, turbulences of different magnitudes and different


frequencies are swept downstream with a speed u,, the convection velocity.
Velocity u, < uo because the friction delays the average speed of the
turbulences relative to the undisturbed flow. Experimental results show
that in the subsonic region

There is a difference in the convection velocities of large turbulence (usually


of low frequency) and small turbulence (usually of high frequency) and
Eq. (4.27a) expresses a mean convection velocity. In the high-supersonic
region (M > 3), experiments show that

For the low-supersonic region the value decreases to between 0.8 and 0.6.
A shear stress exists between the flowing air and the surface. This shear
r, is defined by

where p is the mass density of the air and cf is a friction coefficient that
depends on the smoothness of the surface. For typical aeronautical struc-
tures, experiments of Maestrello15 showed a mean value of cf = 0.0021.
The experiments also showed that the cross and direct PSD of the random
pressure excitation is of the following nature:

Further, it was determined that the cross spectrum depends on the distance
between two points rather than on their absolute location in such a way that
Purchased from American Institute of Aeronautics and Astronautics

46 G. MAYMON
-
where 5 = x , - x2, r] = yl - y2, and fi(O; o ) = 1, fi(O, o ) = 1. Therefore,
P ( o ) is the spectral density of the pressure fluctuations for 77 = 0. A
nondimensional representation of partial test results as obtained by Ref.
z=
15 is shown in Fig. 4.6. For the nondimensionalized test results of Ref. 15,
a closed-form approximation was suggested,

P(")2 . a*O'
' =A, exp(- K, Fw) + A2 exp(- K2Fo) + A3 exp(- K3Fw)
7,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

(4.31)

where

Fig. 4.6 Nondimensionalized P(w) (partial results from Ref. 15).


Purchased from American Institute of Aeronautics and Astronautics

RANDOM FUNCTIONS AND EXCITATION 47

The following flow conditions were selected to plot a dimensional P(o):


M = 0.78, c = 34000 cm/s = 340 m/s = 1115.5 ftls, uo = M . c = 26,520
cm/s, u, = 0 . 8 = ~ ~21,216 cm/s, S = 3.2 cm, and S* = 0.4572 cm = 0.18
in. In Figure 4.7a, P ( o ) as a function of the angular frequency is plotted
on a regular scale. In Fig. 4.7b, it is plotted on a logarithmic scale.
It is interesting to present the PSD of the pressure excitation in terms
of acoustic SPL, a presentation that is more commonly used in acoustic
excitation specifications. These presentations are usually given as a function
of frequency in hertz (cycle/second) and, therefore, P(w) is also presented
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

in Fig. 4 . 7 ~as a function of cycle/second. It should be noted that the areas


under the curves in Figs. 4.7b and 4 . 7 ~should be equal, to yield the same
mean square value for the excitation pressure.
Applying the procedure presented in Appendix B, Fig. 4.8 is obtained for
the pressure excitation in acoustic decibels, for one-third octave frequency
bands, as is usually presented in acoustic specifications. In a usual computa-
tional process, the procedure should be inverted: the decibel levels are
converted into pressure2/hertz, and then to pressure2/(radians/second).
To find the mean square of the pressure excitation, the following integra-
tion is performed:

For the data given for Eq. (4.31) the following is obtained:

resulting in a standard deviation of the pressure fluctuations of


0.0023825 atm.
Equation (4.31) expresses the spectral density of the pressure at a specific
point. In Ref. 14, a cross-spectral density between two different points
(xl, y,) and (x2, y2) is expressed as

where
Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Angular Frequency

Angular Frequency (radhec)

Fig. 4.7 P(o)as a function of frequency.


Purchased from American Institute of Aeronautics and Astronautics

RANDOM FUNCTIONS AND EXCITATION 49

130 1
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Frequency (Hz)
Fig. 4.8 Pressure excitation in acoustic decibels.

and where [and 77 are nondimensional distances along and perpendicular


to the flow, respectively, and a and b are the length and the width of the
rectangular plate subjected to the flow. Equation (4.33) states that the
cross-spectral density of the excitation depends on the distance between
points rather than on the specific location of these points.
Cross correlation between the excitations in different points can be evalu-
ated by using the real part of the Wiener-Khintchine relation

Introducing Eq. (4.33) into Eq. (4.34) yields the following expression for
the correlation function:
Purchased from American Institute of Aeronautics and Astronautics

50 G.MAYMON

where
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Ki and A, are given in Eq. (4.31). The integral Iican be expressed analytically
using the formula

-
- exp((E - KiF)w)[(G - r) sin(G - r)w + ( E - K i F ) cos(G + r)w]
2[(E - Ki F)' + (G + r)']

Using the proper integration limits yields

which can be substituted into Eq. (4.35).


In Fig. 4.9, the autocorrelation functions of the excitation along the flow
(7= 0) and along a line r ) = 5 tan(22.5 deg) for the numerical values given
for Eq. (4.31) are shown, for different values of distances 6 and r). The
value of this function for 5 = r ) = 0 and r = 0 is R,(O, 0; 0) = 5.676205 X
( k g l ~ r n ~which
) ~ , is identical to the mean square of the excitation
calculated by Eq. (4.32).
The maximum point of the autocorrelation functions (Fig. 4.9) corre-
spond to a time delay r, which is equal to

The physical significance of this result is that maximum correlation is ob-


tained by a turbulence at point xl when it is swept downstream with a
velocity u, and reaches point x2 after time 70.
Purchased from American Institute of Aeronautics and Astronautics

RANDOM FUNCTIONS AND EXCITATION 51


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Time Lag (sec)

0 0.0002 0.0004 0.0006 0.0008


Time Lag (sec)
Fig. 4.9 Autocorrelation functions for the turbulent flow.

The results presented in Fig. 4.9 correspond to those obtained by Maes-


trello.15
The model presented [especially Eq. (4.33)] can be used later as a PSD
input for the calculation of the response of rectangular plates to boundary-
layer excitation. An example for this type of analysis can be found in Ref.
17, using the procedures described in Chapter 5, Sec. 111.
Purchased from American Institute of Aeronautics and Astronautics

52 G. MAYMON

V. Summary
Some basic concepts related to random functions and external random
excitations such as mean, mean square, variance, and standard deviation
are presented.
The concepts of autocorrelation, cross correlation, and spectral and cross-
spectral densities and their interrelations, the Wiener-Khintchine equa-
tions, for stationary random functions are presented.
The mean square of a random process is obtained by integration of the
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

PSD function along the frequency axis.


Some practical aspects of the description of the PSD function of an
external excitation are shown.
Some important practical cases of random excitations are shown, includ-
ing a model for the pressure fluctuations in a turbulent boundary layer.
Purchased from American Institute of Aeronautics and Astronautics

Chapter 5

Response of Linear Systems to Stationary


Random Excitations
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Response of a Linear SDOF System

A SSUME that a random force F ( t ) with a mean value m F ( t ) and an


autocorrelation function RF(tl,t2) is acting on a linear SDOF system,
which has an impulse response function h ( t ) , described in Chapter 1. Then

~ ( f =) 1':
- F ( r ) h ( t - 7 )d r = jim ~
- m
(- rt ) h ( r ) d r (5.1)

The definition of the autocorrelation function yields

The mean function m,(t) of the response X can be similarly found as

When F ( t ) is a stationary random function, m~ is not a function of time


and, therefore,

For convenience, w is used in this chapter instead of Q used in Chapter 1.


Using w = Q = 0 in Eq. (1.28b) it is found that

rn,(t) = inr. H(0) = const = rnx (5.5)


Purchased from American Institute of Aeronautics and Astronautics

54 G. MAYMON

For the stationary function F(t), RF(tl, t2) is a function of t2 - tl, so that
RF(tl - 71, t2 - 72) in Eq. (5.2b) is a function only of t2 - t1 - 7, + 7,.
Denoting t2 - tl = T,

Equations (5.5) and (5.6) imply that the response of an SDOF system to
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

stationary random excitations is also stationary.


To find the PSD function of the response X , the Wiener-Khintchine
equations (4.16) are used, yielding

Denoting A = T - 72 + and using Eq. (1.28b),

where [ I* denotes the complex conjugate of [ 1.


Using Eq.(4.12) the autocorrelation function of the response can be
found,

and the mean square of the response is

Equation (5.8) is one of the most important expressions used in the response
analysis of linear SDOF to stationary random excitation. It states that the
PSD of the response can be obtained by multiplying the PSD of the given
input by the square of the absolute value of the complex transfer function,
which is easily calculated for an SDOF system using Eqs. (1.5) and (1.7)
(and replacing 0 with w)

where wo is the resonance frequency of the SDOF system, 5 is its damping


coefficient, and m is its mass.
Purchased from American Institute of Aeronautics and Astronautics

RESPONSE OF LINEAR SYSTEMS 55

It can be verified easily that when units of length, force, and time are
used, the units of the PSD of the excitation force are (f~rce)~/(radian/
second) and the units of the PSD of the response are (length)2/(radian/
second).
For lightly damped SDOF systems, IH(w)l is a function whose values are
very small over most of the frequency axis, with large values only in the
vicinity of the resonance frequency, as shown in Fig. 5.1. Thus, the PSD
of the response has similar features irrespective of the shape of SF(O).
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Even when SF(w) is not a constant, as shown in the figure, most of the
PSD of the response is concentrated in the vicinity of wo. In fact, a very
good approximation is obtained when SF(w) = SF(wO)is used and, thus,

Using Eq. (5.10), the mean square of the response for a constant PSD of
excitation is

For realistic structures where frequencies are always taken positive,

, Transfer Function
Constant Approximate
Excitation

Real Excitation

0 0.5 1 1.5 2
Excitation to Resonance Frequencies Ratio
Fig. 5.1 Transfer function of an SDOF system and PSD of excitation.
Purchased from American Institute of Aeronautics and Astronautics

56 G.MAYMON

It can be verified that

and, thus,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Equation (5.14b) provides an exact result for the mean square value of the
response of an SDOF system subjected to a stationary random excitation
force with constant PSD over the positive frequency axis and a very good
approximation when the PSD of the excitation is not constant. The mean
square response of the system whose transfer function is described in Fig.
5.1 is almost the same for the two different excitations whose PSD function
are also described in this figure.
The PSD of the response of an SDOF system is narrow banded and is
concentrated around the resonance frequency of the system. In Chapter 1
it was shown that SDOF system responses to harmonic excitations are high
when the excitation has a frequency at or near the resonance frequency.
The difference between the response to random excitation and the response
to a harmonic excitation is that the former results in a narrowband random
response (PSD function that is narrow and has high values around the
resonance). This means that the response has frequencies that are close to
the resonance frequency, but the amplitudes are random, with a mean
square value given by Eq. (5.14b). Thus, an SDOF system is a narrowband
filter for both random and harmonic excitations.

11. Response of a Linear MDOF System


The treatment of a linear MDOF system to stationary random excitations
is based on the use of the basic equations of Chapter 4, Sec. I with the
differential equations of this system presented in Chapter 2. The techniques
are almost identical to those presented in the previous section for the
SDOF system.
It was shown that the differential equations of an MDOF system [Eq.
(2.1)] can be written as a set of uncoupled differential equations (2.18),
containing generalized quantities instead of masses and forces. The problem
then can be treated as a set of SDOF systems, solved for the generalized
coordinates q(t), and the final displacement X(t) is obtained using the
transformation (2.12).
In the case of stationary random excitations, the force vector on the
right-hand side of Eq. (2.1) is a vector of random forces. Thus, the general-
ized forces on the right-hand side of Eq. (2.18) are also random. Usually
these forces are defined by their PSD matrix. This matrix is similar to that
Purchased from American Institute of Aeronautics and Astronautics

RESPONSE OF LINEAR SYSTEMS 57

given in Eq. (4.16), where the off-diagonal terms (cross-spectral density)


originate from a possible correlation between pairs of forces.
The MDOF system has a typical normal mode matrix [Eq. (2.7)], which
is repeated here:

[4l =
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Note that the first index refers to the mode number, and the second refers
to the specific location (DOF) in the MDOF system.
Denoting the ith line of the matrix [ 4 ] by L4]J (a row vector),

The physical interpretation of L+,] is the first, second, third, . . . , nth modal
deflection of the jth point of the system.
Unlike the SDOF system, which has only one complex transfer function,
an MDOF system has a diagonal matrix of transformation functions Hl(w),
one diagonal term for each resonance frequency wi,

It is more convenient to exclude M, from the transfer function (5.17) and


to define

which is the analog to H,(o) defined in Eq. (1.8) for the SDOF system.
A generalized cross-spectrum matrix [Sd]of the external excitation matrix
[SF] can be defined,

Here, L41b,lTis a column vector with terms similar to those of Eq. (5.16).
The significance of the denominator M i M j is that after multiplication of
the three matrices in Eq. (5.19), each term i, j in the matrix is divided by
this product.
Purchased from American Institute of Aeronautics and Astronautics

58 G. MAYMON

It is shown in the literature that the spectral density of the jth point of
the MDOF system is given by

The matrices in Eq. (5.20) are defined in Eqs. (5.16), (5.18), and (5.19).
Equation (5.20) is the MJOF analog to Eq. (5.8) of the SDOF system.
The inner multiplication [HI [S,] [H*] is the analog to [HI [SF] [H*] and
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

the multiplication in L+j] and its transpose originates from the transforma-
tion (2.12).
To solve specific problems, the designer is mainly interested in the mean
square value E[X,2] of the response at point j and in its standard deviation.
The mean square of the response is obtained by integrating Sx,(w) over
the frequency range; thus,

When the excitation has a zero mean (which is a frequent case in random
vibration analysis of practical systems), E [ X f ] is equal to the variance of
X , and the standard deviation of the response is

Although Eq. (5.21) appears complicated, in effect this is not the case.
When the resonance frequencies and mode shapes are known, and when
the excitation is given, i.e., by specifications, all of the matrices in this
equation are known. If the PSD of the excitation does not depend on the
frequency, it can be taken out of the integral, and a much simpler definite
integral can be solved. Expressions for important integrals of frequent cases
of practical importance can be found in Appendix A. If the PSD of the
excitation is frequency dependent, a numerical integration can be used.
Auto- and cross correlations can also be found, using the Wiener-
Khintchine formulas (4.12). As these quantities are of lesser importance
to the practical user, their expressions are not presented here, and the
interested reader can find them in the literature, e.g., Ref. 11.
More features of the solution presented in Eq. (5.20) are cleared from
the following numerical example.

Example 5.1
A two-DOF system is described in Fig. 5.2. In this system, there are two
masses m land m2 and two springs kl and k2 whose values are selected so
that the resonance frequency of each mass on its own springs is o, for the
first mass and o, for the second mass. The damping coefficient of ml on
the spring kl is 5, and that of the mass m2 on the spring k2 is ls.A random
force f(t) with one-sided white noise (constant value over the positive
Purchased from American Institute of Aeronautics and Astronautics

RESPONSE OF LINEAR SYSTEMS 59


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

(t> X2 (t)
Fig. 5.2 Two-DOF system.

frequency axis) spectral density of magnitude Soacts on the first mass. This
is a typical model of a primary randomly excited system, which supports a
secondary system. Responses x, (t) and x2(t) are selected as the generalized
coordinates q, and q2. The differential equation of the system is

The following numerical values are selected: ml = 1 kg s2/cm, m2 = 0.1


kg s2/cm, wp = w, = 125 radls, l p= 5, = 0.01, and So = 550 (kg)'/(rad/s).
Solving the characteristic Eq. (2.6) yields

where the modal displacements (the normal modes) of the first mass were
arbitrarily selected as a unity. The normal modes matrix is, therefore,
Purchased from American Institute of Aeronautics and Astronautics

60 G. MAYMON

The generalized mass matrix is [Eq. (2.8)]

The modal damping coefficients are obtained by solving


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

therefore,

Note that the modal damping coefficients are not identical to the physical
damping coefficients of the primary and the secondary systems.
The excitation spectral density matrix is

and by use of Eq. (5.19)

The spectral density of the response xl of mass ml is

s x , = L41I[ @ 4 l [S,(o)l [H(o)l*L41 IT


Purchased from American Institute of Aeronautics and Astronautics

RESPONSE OF LINEAR SYSTEMS

and for the mass m2


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Note that in both Eqs. (5.30) and (5.31) the PSD of the response depends
on three terms. The first one contains IHl(w)I2and is due to the contribution
of the first resonance to the vibration of the selected point. The second
term contains IH3(w)I2and represents the contribution of the second reso-
nance. The third term contains two mixed products of the transfer functions
and their conjugates and is the result of the interaction between the two
resonances at the selected point. For a system with more than two DOFs,
n direct terms and n ( n - 1) mixed terms exist, the latter~epresengingthe
- - between pairs of modes. Using the identity HjHZ + HYH, =
interaction
2 Re(H,H;), the number of the mixed terms can be reduced to n(n - 1)/2.
Integrating Eq. (5.31) over the frequency range, and using the preceding
identity and the integrals presented in Appendix A, yields

therefore,

The same process is repeated for Eq. (5.32) to yield

E(xg) = 0.10127 + 0.02873 - 0.0001678 = 0.129547 cm2

Therefore, the two masses vibrate so that the root mean square of the
displacement of the first and second masses are 0.1065 and 0.3603 cm, re-
spectively.
Purchased from American Institute of Aeronautics and Astronautics

62 G. MAYMON

In the preceding solutions for the mean square value of the responses,
the three terms described earlier are presented separately. It can be seen
that the contribution of the first resonance is larger than that of the second
one. For modal dampings that are similar (though not equal) this is always
the case, because of the term wq in the denominator of Eq. (Al). The
contribution of the (third) interaction term is much smaller. This happens
when the resonance frequencies are well separated from each other, due
to the term ( w f - ~ 7 in )Eq.~(A6).
This observation can be used to find an approximate solution for an
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

MDOF system with well-separated resonance frequencies, by neglecting


the interaction terms between pairs of resonances. The procedure, which
can be easily proved and is not included here, takes the following steps:
1)Write the matrix S4(w)by using Eqs. (2.8), (5.15), and (5.19). Assuming
that the mode shape matrix and the mass matrix are given by numerical
values, the result matrix is a function of the SF matrix functions. Do not
forget to divide eachterm by MiM,.
2) Build a matrix S4(w), which contains only the diagonal terms of S4.
These terms contain expression of the known SqSj(w)functions.
3) For the first diagonal term, calculate a numerical value S41,1by inserting
o = wl in the SFi,,(w)functions. For the second diagonal term, calculate
a numerical value S42,2 by inserting w = w2 in the SF,,,(w)functions, etc.
4) The expected mean value of the displacement response XI is then

5) For any point j,

The procedure is shown schematically in Fig. 5.3.

111. Response of a Linear Structure to Stationary Random Excitations


In Chapter 3 it was shown that a continuous structure can be treated as a
collection of an infinite number of SDOF systems, provided that generalized
masses, modal damping coefficients, and generalized forces replace the
regular quantities. It was also assumed that resonance frequencies and
modal shapes of the N first resonances of the infinite number possible
can be evaluated, i.e., by a finite element program, and the structural
displacement can be evaluated approximately with these first N modal
shapes + i ( x ) and the generalized coordinates qi(t).Therefore, a solution
Purchased from American Institute of Aeronautics and Astronautics

RESPONSE OF LINEAR SYSTEMS

Data
Masses, frequencies,
mode shapes,
modal darnpings,
excitation PSD matrixs SF
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Calculate generalized
masses matrix [MI

Write expressions for Sg


using (2.8), (5.15), (5.19).
Remember to &vide to the
generalized masses

Calculate numerical values


of the diagonal terms of Sg

Calculate mean square


values using (5.32)

Fig. 5.3 Procedure for an approximate response solution of an MDOF system.

of a continuous system is an analog to an MDOF system, where

Suppose that a continuous structure is excited by random forces that have


the cross-spectral density function Sq(xl, x2, o), which is an analog to the
[SF(o)] in Eq. (5.19) for the MDOF system. This cross-spectral density
function can be evaluated in terms of the normal modes and a set of
coefficients SQjQkso that
Purchased from American Institute of Aeronautics and Astronautics

64 G.MAYMON

It can be shown that

Equation (5.36a) is the continuous-structure equivalent to Eq. (5.19) of the


MDOF system, where each term j, k is an analog to the compatible term
in the matrix [S$(o)] of the latter system. In the numerator of Eq. (5.36a),
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

a double integration over the whole structure is performed. Thus, in a one-


dimensional structure, say, a beam of length L , two integrations for 0 5
xl 5 L , 0 5 x2 5 L are included. For a two-dimensional structure, say, a
rectangular plate of length a and width b, two integrals for 0 5 xl 5 a, 0
5 x2 5 a, and two for 0 5 yl 5 b, 0 5 y2 5 b are included. Thus, for a
one-dimensional structure (beam, frame)

and for a rectangular plate,

These equations can be written in a nondimensional form for r = xlL for


a beam, or 5 = xla and 7 = y l b for a plate; thus,

for a beam, and

for a plate. The generalized masses can be written as


Purchased from American Institute of Aeronautics and Astronautics

RESPONSE OF LINEAR SYSTEMS 65

for the beam and the plate, respectively, where r n is~ the mass per unit
length of a beam and r n is ~ the mass per unit area of a plate. Both may
be a function of the location on the relevant structure.
Sometimes the spectral density of the excitation may be separable into
a term that depends on the frequency and a term that depends on the
location for a beam and a plate, respectively,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

In this case, So(o) can be taken out of the integrals in Eqs. (5.36d) and
(5.36e).
All of the terms of Eqs. (5.36) can be easily calculated for a given structure
once the mode shapes and the cross-spectral function of the excitation are
given. The use of a finite element program results in mode shapes and
generalized masses. In most of the commercially available finite elements
computer codes, the modes are normalized in such a way that all of the
generalized masses are equal to unity and, thus, the matrix of the generalized
masses is the unit diagonal matrix [ I ] .
The cross-spectral density of the response of the structure is given by an
expression analog to Eq. (5.20) of the MDOF system. For a one-dimensional
structure, it is

Equation (5.39) is the continuous system equivalent of Eq. (5.20) of the


MDOF system. It is the cross-spectral density function between the points
and 6on the structure, due to all of the modes, i, k = 1, . . . , N. The
transfer functions in Eq. (5.39a) are those given in Eq. (5.18). If & Z t2,
el
the cross-spectral density between and t2is obtained. If t1= & = 6, the
spectral density at point 5 is obtained

For a two-dimensional structure like a plate,

which is the cross-spectral density between two points (el, 77,) and (t2, q2).
If 51 = 6 = 5 and 771 = rl2 = 7, the spectral density of the response at
Purchased from American Institute of Aeronautics and Astronautics

66 G. MAYMON

point (6, 7) is obtained

The mean square value of the response is obtained by integrating the


spectral density of the response over the frequency range. It should be
noted that in many references the spectral density of the input is assumed
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

to have components in both negative and positive frequencies. This is the


outcome of the Wiener-Khintchine formulas (4.12), where S, is obtained
from the autocorrelation function R,, which exists for both negative and
positive time delays T.
In cases where the spectral density is given by specifications, standards,
or experimental measurements, it is one sided and, therefore, the values
for negative frequencies are identically zero. In the particular engineering
cases treated here, the spectral density is always taken as a one-sided
function, for positive values of w only. Therefore, using Eq. (5.39b), the
mean square of the response of a one-dimensional structure, such as a
beam, subjected to one-sided spectral density is

A very similar expression is obtained for a two-dimensional structure, where


the only difference is that the mode shapes are functions of both 6 and 7.
Sometimes it is of interest to calculate the spectral density of the time
derivative of the response, for instance, the spectral density of the velocity.
It can be shown that

and, therefore, the mean square of the time derivative of the response is

In Eq. (5.41) two kinds of terms exist: integrals of


-
H,?(w) . Hi(w) = IHj(w)12 when j =k (5.44a)

and integrals of

B,*(w) - Hk(o) + H,(w) - H,*(w) when j # k (5.44b)


Purchased from American Institute of Aeronautics and Astronautics

RESPONSE OF LINEAR SYSTEMS 67

which are similar to the expressions (5.30) and (5.31) in the example of
the two-DOF system. Thus, direct terms (contributed from each resonance)
and cross terms (contributed from the interaction of pairs of resonances)
exist in the response at any given location.
Because of Eqs. (5.44), Eq. (5.41) can be rewritten as
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Using the identity H,HZ + H,*Hk = 2 Re[HjHz], the last equation takes
the form

where the first summation term contains N terms, which are the outcome of
the N resonances and the second summation term contains N(N - 1)12,
which are the result of intermodal interaction. Thus, Eq. (5.45b) contains
N(N + 1)/2 terms, whereas Eq. (5.41) contains I\IL terms. A similar expres-
sion for the expected value of the velocity can be obtained by multiplying
the integrands in Eq. (5.45b) by w2.
As in the MDOF system, a quick practical approximation for cases in
which the natural frequencies are well separated and the modal dampings
are small can be obtained, for the cases where the spectral density does
not vary strongly in the vicinity of the resonance frequencies. In these cases,
the interaction between the well-separated resonances is negligible. Also
the response has high values around the resonances and negligible values
below and above these frequencies. Each resonance is then treated as an
SDOF system, and the contribution of the jth mode to the mean square
of the response is

The total mean square is obtained by summing the contribution of N modes.


When the external excitation has a one-sided constant value spectral
density function So(@),which does not depend on the location on the
structure, Eq. (5.46a) takes the form
Purchased from American Institute of Aeronautics and Astronautics

68 G. MAYMON

It can be seen that the contribution of the jth mode to the mean square
of the response is proportional to my3; thus, for well-separated resonances
the main contribution to the mean square of the response comes from the
first resonance. Therefore, in the cases treated by this approximation, a
reasonable engineering result can be obtained using only one DOF. This
conclusion is less accurate for the derivative process (velocity response),
because in this case
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

and the contribution is proportional to o7'. The approximation is even less


accurate for the second derivative process (acceleration). Therefore, it is
usual to see engineering structural systems that have relatively high-
frequency acceleration responses, whereas the displacement response is
concentrated in the low-frequency range.
In Chapter 3, Sec. IV, the concept of stress modes *f')(t) was introduced,
where j is the mode number and (i) is an index that indicates which stress
is being calculated. Using Eq. (3.39), expressions analog to Eqs. (5.39a),
(5.39b), (5.40a), (5.40b), (5.41), (5.43), (5.45a), (5.45b), and (5.46a-5.46~)
can be obtained for the stress response and the stress time derivative, by
changing 4i([) with *f')([). For instance, Eq. (5.46b) becomes

for the contribution of the jth resonance to the mean square of the stress
(i). Similarly (5.46~)becomes

Thus, all of the procedures, expressions, and numerical programs used for
the computation of the displacement response of the structure can be used
for the calculation of the stress response, using the concept of stress modes.
These stress modes can be calculated using the techniques described in
Chapter 3, Sec. IV.

Numerical Example
A simply supported steel beam of length L with a uniform rectangular
cross section of width b and height h is subjected to a stationary random
excitation force per unit length whose PSD function S, is a function of only
the angular frequency w, as shown in Fig. 5.4. The root mean square value
of the displacements and the bending stresses at the mid- and quarter-
Purchased from American Institute of Aeronautics and Astronautics

RESPONSE OF LINEAR SYSTEMS


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Angular Frequency
Fig. 5.4 PSD of the excitation.

length of the beam is required. For simplicity, only three normal modes
are used.
The following numerical values are taken: b = 2 cm, h = 1 cm, L = 30
cm, density p = 7.9592 X kg s2/cm4, Young modulus = 2.1 X lo6
kg/cm2, and modal dampings = 0.015, l2= 0.02, and L3 = 0.02.
From the geometry, the cross-section area is A = bh = 2 cm2, and the
area moment of inertia is I = 1/12 bh3 = 0.16667 cm4.The first three frequen-
cies and mode shapes of a simply supported beam are given by

o,=
L2
/$ = 16'26.09 radls; @I(() = sin(n0

The generalized masses are given by (for example, see Ref. 18)

M, = yo
p A L sin2 (jn() d t = M p A L

Ml = M2 = M3 = 2.38776 - kg s2/cm
Purchased from American Institute of Aeronautics and Astronautics

70 G. MAYMON

To calculate SQ,Q,, Eq. (5.36d) is used. For this purpose, the following
integrals are required:

1;sin ( n n d( = -;
'A
1: sin (2770 d c = O; J1sin ( 3 n d~c = -
2
3n

and, therefore,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where S,(w) is given by Fig. 5.4. Note that all terms connected to the second
resonance are zero. This is because the second mode is antisymmetric while
the excitation is uniform (and, therefore, symmetric) along the beam. No
response in the second mode and no interaction between the second modes
and the first or the third modes are expected.
Equation (5.46a) is used for the computation of the mean square values
of the response. To use the equation, three values from Fig. 5.4 are required,

and the values of the modal displacement at the midbeam and quarter-
beam are also required

Substituting the numerical values in Eq. (5.46b) yields

Therefore, the root mean square (rms) value of the displacement at the
center of the beam is 0.5216 cm and at the quarter-beam is 0.1360 cm. To
calculate the mean square of the bending stresses, the stress modes for
bending of a beam are required. One of these has already been calculated
Purchased from American Institute of Aeronautics and Astronautics

RESPONSE OF LINEAR SYSTEMS 71

in Example 1 of Chapter 3, Sec. IV, and is repeated here, together with


the second and third bending stress modes;

q@end) = 1 ' r2Eh sin (23-5)


1
2L2
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

.$bald) = 9 . v 2Eh sin (323-5)


2L2

Substituting the numerical value yields

Substituting these values into Eq. (5.47a) yields

The rms of the bending stresses are 6088 kg/cm2 and 4304 kg/cm2, respec-
tively. It should be noted that the rms values are not the maximum value
of the response. If, for instance, the excitation has a Gaussian (normal)
distribution, the response is also Gaussian and values of 3 u may occur in
99.73% of the applied time.
The major contribution to the response is from the first mode. As was
earlier mentioned in this section, the second mode does not participate in
the response. Note that the contribution of the third mode to the stresses
is higher relative to its contribution to the displacements. This is due to the
higher curvatures of the third mode, which result in higher bending stresses.

IV. Summary
In the evaluation of the expressions in this chapter, linearity is assumed
for the SDOF, MDOF, and continuous systems, and the excitation is as-
sumed to be stationary. Therefore, one can use the formulation presented
for continuous, stationary random vibration, but not for the response to
transient excitations.
Expressions for the computation of the mean square values of the dis-
placements of these systems are also presented. When the excitation has
zero mean, the mean square values are equal to the variance of the response.
Purchased from American Institute of Aeronautics and Astronautics

72 G. MAYMON

Approximate solutions are presented for MDOF and continuous systems


that have well-separated resonance frequencies. These approximations are
obtained by neglecting the effects of interaction between modes.
For continuous systems, the displacement response expressions can be
used for stress response, by replacing the normal modes in the relevant
equations with the stress modes.
Stress modes can be calculated by a commercial finite element program,
using the procedure described in Chapter 3, Sec. I.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494
Purchased from American Institute of Aeronautics and Astronautics

Chapter 6

Nonlinear Single-Degree-of-Freedom and


Multiple-Degree-of-Freedom Systems
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Introduction

A NONLINEAR structure behaves differently than a linear one. Factors


such as geometrical nonlinearity (hardening or softening) of a struc-
ture due to geometrical changes when large deflections are introduced,
material nonlinearity, e.g., plasticity, and changes in damping coefficient
as a function of the structural response may increase or decrease the defor-
mation and, thus, the stress distribution in a structure, as compared to the
linear model. In many realistic structures, a linear model may result in a
conservative design, both in static and dynamic analysis. Therefore, the use
of nonlinear analysis may provide more efficient, less conservative designs.
There are many methods that are used for the nonlinear treatment of a
structure. All of the commercially available large finite element codes en-
able the calculation of the static behavior of a nonlinear structure. The
nonlinear effects that can be handled by these programs include geometrical
nonlinearity (large deflections and stiffening effects) and material nonline-
arity.
Fewer computational tools are available for the nonlinear vibration analy-
sis of realistic structures. Although many analytical methods for such analy-
ses exist, e.g., Refs. 19-21, the implementation of these methods for compu-
tation of practical structures is not straightforward. Usually, due to design
considerations, good engineering practice is to avoid the nonlinear material
range in vibration. Nevertheless, elastic (geometric) nonlinearity may be
included in the vibration analysis to obtain significantly less conservative
designs. Therefore, this chapter is limited to geometrical nonlinearity.
A classical example for an elastic nonlinear system is a plate (either
supported or clamped at all edges), where the supports cannot move toward
each other when the plate is deformed by a perpendicular load. As the
deformation increases, the length of the mean surface of the plate increases
and membrane stresses are generated, in addition to the classically com-
puted bending stresses. Because of these membrane stresses, the lateral
deflection of the plate is smaller compared to the lateral deformation when
the supports can move. Significantly less lateral deflection and, thus, less
bending stresses are obtained.
To familiarize the reader with the major phenomena of geometrical
nonlinearity of a structure, a description of an SDOF oscillator is described
Purchased from American Institute of Aeronautics and Astronautics

74 G. MAYMON

in Sec. 11. In addition, the practical treatment of an MDOF system, repre-


senting a continuous structure, is described in Sec. 111.

11. Nonlinear Behavior of an SDOF Oscillator


In a linear SDOF system, the relation between the restoring force F and
the displacement q is
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where a is the linear rigidity of the system.


In most conventional cases, the response of the structure to loads in
one direction is antisymmetric to the behavior in the opposite direction.
Therefore, the nonlinear behavior of the restoring force can be described
by an odd series

It can be shown that in many cases the use of a cubic equation [only two
terms of Eq. (6.2a)l can accurately approximate the nonlinear behavior.
An oscillator with a cubic force-displacement relationship

is called a Duffing oscillator. The behavior of a Duffing SDOF oscillator


of mass m, subjected to an excitation force F(t) is

where wo is the resonance frequency of the linear oscillator, P is a coefficient


of nonlinearity, and l is the damping coefficient in small amplitudes. When
the oscillator is excited by an harmonic force F = Fo cos ot, the steady-
state response will be harmonic, with frequency o and amplitude A. It was
shown19 that the following relation exists:

For a given set of Fo, M, 5, wo, and p, the relation between A and o is
described in Fig. 6.1. Equation (6.4) degenerates to the classical expression

for linear systems where /3 = 0.


The nonlinear response curve (6.4) has an overhang to the right when
j3 > 0 (hardening), as shown in Fig. 6.1, and to the left when /3 < 0
(softening). The turning point B lies on a line, which is called the backbone
curve. For A = 0 this curve intersects the frequency axis at o = o,d =
Purchased from American Institute of Aeronautics and Astronautics

NONLINEAR SDOF AND MDOF SYSTEMS 75


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

0.7 0.8 0.9 1 1.1 1.2 1.3


Relative Frequency
Fig. 6.1 Nonlinear response of an SDOF oscillator.

wo m.The backbone curve is given by


For a damping coefficient that is small relative to 1, Eq. (6.6a) degener-
ates into

It can be seen in Fig. 6.1 that the resonance frequency in higher amplitudes
is higher (in the case of hardening system) than the resonance frequency
of the linear (small-amplitude) response, due to the higher rigidity of such
a system.
It is interesting to see that there is a region of o for which three solutions
Purchased from American Institute of Aeronautics and Astronautics

76 G.MAYMON

of A exist for a given excitation frequency. The stability of these three


solutions was checked.19 Experimental work showed a jump behavior be-
tween the points. Thus, when a constant amplitude force, with a frequency
that changes from zero upward, is applied, the amplitude of the response
follows along the curve ODAB. The response jumps to B1 and follows the
curve BICl. If the excitation frequency is decreased, the response follows
the curve CIBIDl and then jumps to point D and fo1low.s the curve DO.
Many characteristics of the oscillator can be deduced from this type of
experimental procedure, some of these were described in Refs. 22 and 23.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

When this type of a system is excited by a wideband random excitation,


its response contains components in many frequencies, and the jump phe-
nomena, which exist when the excitation is harmonic, disappear. In some
cases of narrowband random excitation, the jump can still be observed
because the basic characteristic of the response to narrowband excitation
is very similar to a harmonic excitation, which is a limit case of a nar-
rowband excitation.
Sometimes a continuous system can be approximated by an SDOF sys-
tem. This is the case when most of the mass of the structure is concentrated
around a single location and its elasticity is contributed by a well-defined
elastic part. It is advantageous to create this type of SDOF approximation
in the early stages of a particular design, due to the simplicity of the solution,
which can enhance the understanding of the behavior of the continuous
system and point out some important parametric effects. In the following
example, a continuous structure is approximated by an equivalent SDOF
Duffing oscillator, and response curves for different excitation levels are
calculated and compared to linear analysis results.
A uniform beam is simply supported on immovable supports as shown
in Fig. 6.2a. The dimensions and material properties are also shown in the
figure. The beam is replaced by an SDOF system described in Fig. 6.2b.
In the equivalent oscillator, an equivalent concentrated mass mll is located
at the center of the beam, and the springs are considered to have no mass.

Fig. 6.2 Beam with immovable supports and its SDOF equivalent: L = 60 cm,
b = 8 cm, h = 0.5 cm, I = 0.08333 cm4, E = 2.1 x lo6 klcm? p = 7.959 x
(specific weight 7.8), & = 0.02, W, = total weight = 1.872 kg, and M, = total
mass = 1.902 x kg &em.
Purchased from American Institute of Aeronautics and Astronautics

NONLINEAR SDOF AND MDOF SYSTEMS 77

The static linear relationship between the midspan force Fl and the
midspan deflection r], is

and k l l can be calculated from the beam formula^.'^ With the given data
one obtains
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The equivalent mass is calculated so that the linear resonance frequency


of the equivalent oscillator coincides with the beam first resonance fre-
quency, also calculated by known formulas, e.g., see Ref. 1,

Therefore,

It should be noted that the generalized mass of the beam is not necessarily
equal to the equivalent mass

and the difference between M and mll is 1.4%.


The static nonlinear equation for the midspan deflection is

This equation can be used to calculate the coefficient of nonlinearity of


the system. Although the behavior of the displacement of the mass at the
center of the beam as a function of the static force Fl can be solved
analytically, a finite element nonlinear module of the ANSYS program is
used. This is done to demonstrate a solution process that is applicable to
more general problems. The nonlinear module of the ANSYS finite element
program, where large deflection and stress stiffening effects are taken into
account, is used. The beam is divided into 24 beam elements and is numeri-
cally loaded by a midspan 300-kg force. Results of deflections as well as
the results of a linear analysis are described in Table 6.1.

Table 6.1 Linear and nonlinear midspan deflections

Nonlinear q,
F, kg Linear 7, cm, analytic Linear 7, cm, ANSYS cm, ANSYS

300 7.71459 7.7146 0.8822


Purchased from American Institute of Aeronautics and Astronautics

78 G. MAYMON

The nonlinear analysis result is noted to be 8.74 times smaller than the
linearly calculated deflection. This indicates a significant stiffening effect
for this structure.
Substituting q and F from Table 6.1 and value of kll given earlier into
Eq. (6.8) yields the value of bll (kg/cm3)

Therefore,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The equation of motion of the equivalent SDOF oscillator is [see Eq. (6.3)]

125 150 175 200 225 250 275 300 325


Frequency (radlsec)
Fig. 6.3 Nonlinear and linear response of the equivalent oscillator.
Purchased from American Institute of Aeronautics and Astronautics

NONLINEAR SDOF AND MDOF SYSTEMS 79

Assume Fl = Fo sin ot, then solutions to Eq. (6.9) can be plotted for
different values of harmonic excitation forces of amplitudes Fo. This is
shown in Fig. 6.3. The difference between ooand odis neglected. It should
be noted that even for deflections that are smaller than the beam height
( h = 0.5 cm), significant reductions in the maximum response are obtained.

111. Nonlinear Coefficients of a Structure


To solve a nonlinear problem, a knowledge of the nonlinear coefficients
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

of the structure is required. When a discrete system of several masses and


springs are given, the nonlinearity of the springs and the nonlinearity of
damping (if such exists) is usually given as data. This is not the case for a
nonlinear structure. The structural geometrical nonlinearity is an outcome
of the structural configuration and its boundary conditions. Although ana-
lytical expressions that describe nonlinear behavior may exist for some
simple structural elements, the practical structures encountered in engi-
neering practice must be solved through a numerical algorithm that usually
requires discretization of the structure. Also, it should be kept in mind that
a nonlinear system cannot be solved by a superposition of several simpler
nonlinear cases, as the superposition principle exists only for linear systems.
Therefore, the method for calculating the coefficients of nonlinearity is
to solve static cases using a nonlinear finite element code and to extract
nonlinear coefficients from these solutions. The use of finite element pro-
grams for structural analysis is practically a discretization of a continuous
system (infinite number of DOFs) into an MDOF system. The method for
doing so is better demonstrated by a numerical analysis of the simply
supported beam described in Fig. 6.2, this time without modifying it into
an SDOF system but into an MDOF system. The designer has to decide
how many generalized coordinates will participate in the analysis. This
decision must be based on logical considerations that depend on the nature
of the problem and the past experience of the designer. Often symmetric
conditions of the structure and the loads can significantly decrease the
required number of generalized coordinates. If the behavior of only a few
points on the structure is important, the response of only these points
should be selected as general coordinates.
For the demonstrated example, three general coordinates q l , q2, and q3
at quarter span, midspan, and three-quarter span, respectively, were se-
lected. The beam is divided into 24 beam elements, as shown in Fig. 6.4.

Fig. 6.4 Discretization of a continuous beam into three-DOF system.


Purchased from American Institute of Aeronautics and Astronautics

80 G. MAYMON

First the linear case is treated,

or more specifically,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The beam is load'ed with LF1 0 01T, and Lql q2 q3] is calculcated with
a linear finite element code. Substitution of this vector of results into Eq.
(6.10) yields three equations with nine kij unknowns. The beam is then
loaded with 10 F2 0IT to obtain three more equations, and then loaded
with LO 0 F3ITto obtain another three equations. Thus, nine equations
enable a solution for the nine kij. In Fig. 6.5 the numerical values of the
loads and the linear deflections are shown.
The nine equations are

Solution of these equations yields

Although the demonstrated problem is symmetric and can be solved with


only two general coordinates, three general coordinates are shown to dem-
Purchased from American Institute of Aeronautics and Astronautics

NONLINEAR SDOF AND MDOF SYSTEMS

300 0 0 Load
A I 1
I

4.3395 5.3038 3.3751 Linear


0.6886 0.6295 0.3293 Non-linear
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

0 300 0 Load
I I I

5.3038 7.7146 5.3038 Linear


0.5395 0.8822 0.5395 Non-linear

0 0 300 Load
1 I I

3.3751 5.3038 4.3395 Linear


0.3293 0.6295 0.6886 Non-linear
Fig. 6.5 Linear and nonlinear deflection for three loading cases.

onstrate a more general case. Also, it is not necessary to use loading vectors
that contain zero values. This simplifies the set of algebraic equations but
this is not necessarily required in a general computerized procedure.
Now, assume that the nonlinear behavior of the structure can be formu-
lated by

or more explicitly,

Using the same three loadings as for the linear case, but using the nonlinear
module of the ANSYS program, one obtains the deflections for the nonlin-
ear case. These are also shown in Fig. 6.5. Substituting these deflections
Purchased from American Institute of Aeronautics and Astronautics

82 G. MAYMON

into Eq. (6.13) together with the values of the linear rigidity given in Eq.
(6.12) the following is obtained:
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

It should be emphasized that there is no connection between the number


of finite elements selected for the solution of the problem (24 in this
example) and the selected number of generalized coordinates. The number
of elements selected is influenced by the requirements imposed on the
static finite element solution, whereas the selected number of generalized
coordinates is influenced by a prior estimation of the number of modes that
are expected to significantly participate in the dynamic response analysis.

IV. Summary
The behavior of a nonlinear SDOF oscillator is described.
Geometric antisymmetric nonlinearity is assumed. In particular, a Duffing
oscillator is treated. As a result of this nonlinearity, the natural frequency
of the oscillator is amplitude dependent, with an increased frequency for
a positive coefficient of nonlinearity and a decreased frequency for a nega-
tive coefficient.
A jump behavior may exist in a geometric nonlinear oscillator.
A practical method for calculation of the nonlinear coefficients of a
practical structure using a finite element program is described and demon-
strated for a simple structure, which is represented by a three-DOF system.
Purchased from American Institute of Aeronautics and Astronautics

Chapter 7

Statistical Linearization Method


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Statistical Linearization Method for an MDOF System

A WIDELY used methodz4for the treatment of nonlinear systems is


the statistical linearization method. As the analysis of a realistic struc-
ture is usually performed by discretization of the continuous system, the
analysis procedure is demonstrated for an MDOF system.
The general equation of motion of a nonlinear MDOF system can be
written as

The functions @ are nonlinear functions of the vector of the generalized


coordinates { q ) and its first and second time derivatives.
There are only a few cases where an analytical solution of Eq. (7.1) can
be obtained knowing a. Several methods for a solution were suggested in
the literat~re.'~ Some of them are based on perturbations, energy balance,
and slowly varying parameters. Nevertheless, these techniques fail in many
practical cases, or develop into very complicated procedures. The concept
of statistical linearization permits a systematic approach, which can be
mechanized into a general approximate solution. According to this method,
the nonlinear equations are replaced by equivalent linear equations, which
are selected in such a way as to minimize the difference between these
and the original equations. When an equivalent set of linear equations is
obtained, they are solved by known techniques of linear equations.
The equivalent linear equations are defined as

where me, c,, and k, express additional equivalent mass, damping, and
rigidity, respectively. The difference between the original Eq. (7.1) and the
equivalent Eq. (7.2) gives a difference vector {E), which is

The mean square of the difference E is minimized

E { s T . E ) = minimum (7.4a)
Purchased from American Institute of Aeronautics and Astronautics

G.MAYMON

E{E: + E$ + . . . + E:} = minimum (7.4b)

where ci are the elements of the vector {F). Equation (7.4b) can be written as

2 DT = minimum
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

i=l

where D; = E ( E ~ )i, = 1, 2, . n.
Denoting m;, c;, and k;, the elements of [m,],[c,],and [k,],respectively,
and denoting Qi the elements of {+I, D: takes the form

Minimization of Df is obtained when

Introducing Eq. (7.5) into Eq. (7.6) yields the following set of equationsz4
after some algebraic manipulations:

where qT = Lq, q, ql and m$, c$, and &* are the ith row of the matrices
[m,], [c,], and [k,], respectively. Equation (7.7) is a set of equations for
me.11
9
ce.
I]
7
andk'.(1'
The equations for the equivalent system can be simplified when the
excitation is Gaussian, which is a frequent assumption in the analysis of
the response of structures to random excitation. For linear systems, the
response to Gaussian excitation is also Gaussian. Therefore, an approxima-
tion of Gaussian response for a nonlinear system excited by Gaussian
Purchased from American Institute of Aeronautics and Astronautics

STATISTICAL LINEARIZATION METHOD 85

excitation is quite logical. It was shown in the literaturez4that the following


formula exists for a Gaussian vector variable r]:

E { f( d . r ] ) = E{r]. v T ) E{Vf
. (7)) (7.8)
where
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

When (4) in Eq. (7.7) is assumed to be Gaussian, comparison of Eqs. (7.7)


and (7.8) leads to

where

and similar relations are obtained for d @ J d q and dQiIaq.


Purchased from American Institute of Aeronautics and Astronautics

86 G. MAYMON

Comparison of Eq. (7.9) to Eq. (7.7) yields


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

In the evaluation of the former equations, it was assumedz4that the nonlin-


ear functions are antisymmetric

Therefore, the response, which was assumed to be Gaussian, has a zero


mean, which also requires the Gaussian excitation to have a zero mean.
This is a reasonable assumption when analysis of the response of structures
to external Gaussian excitation is performed. When the excitation does not
have a zero mean, it must be separated into a zero mean excitation and a
static bias. the treatment of the latter is not included in the present dis-
cussion.
Equation (7.10) includes the expected values of the derivatives of the
nonlinear functions @withrespect to the expected values of the generalized
coordinate vector {q} and its time derivatives. These expected values are
not known before the solution is determined. The expression that relates
the Gaussian excitation to the expected values of the response was already
presented in Chapter 5, Sec. 11. If the excitation forces {Q(t)}in Eq. (7.1)
are random Gaussian and have a spectral density matrix SQ(w),the weighted
generalized spectral density matrix of the generalized forces is given by
[see Eq. (5.19)]

and the spectral density of the response is given by [see Eq. (5.20)]

Here H ( o ) rather than H ( w ) is used, because during the solution process


the generalized process the generalized masses are changed. The variance
of the generalized coordinate q, is given by the integral of the relevant
term of the response spectral density over the frequency range
Purchased from American Institute of Aeronautics and Astronautics

STATISTICAL LINEARIZATION METHOD 87

Equations (7.10), (7.12), and (7.13) provide expressions for the computation
of the expected values of the response {q} of the equivalent linear system,
which must be solved iteratively. First, the nonlinear effects are neglected
and the variance of the response of the linear system is calculated. Then,
m&,c& and k;, are calculated with Eq. (7.14) and added to the equivalent
equation of motion. New values of the variance are computed, and a new
set of m;, c;, and k&,are computed. The process is repeated until a conver-
gence of the variance is obtained, according to the required accuracy.
Continuous structures are usually defined by their normal modes and
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

can be treated as MDOF systems, where the number of DOFs taken into
account is the number of modes involved in the response process and where
the generalized masses, rigidities, and modal damping coefficients are used.
The integral in Eq. (7.14) takes the form

When S+(w) is constant, integrals such as Eq. (7.15) can be performed using
the formulas of Appendix A. When S+(w) can be represented as filtered
white noise, Appendix A can also be used. If S+(w) is a rational function
of w, a method of solution for the integral is described in Appendix B of
Ref. 24. Nevertheless, numerical integration may be required for complex
practical structures. It should be borne in mind that S+(w) is not the original
matrix of the spectral density of the excitation, but a matrix obtained by
a suitable process, where the original matrix is weighted by the normal
modes, as seen in Eq. (7.12).
During the iterative process of solution of the equivalent linear equations,
modes are recomputed in every iteration and, therefore, the generalized
masses, modal frequencies, and modal damping coefficients are changed
during the process.

11. Nonlinear Response of an SDOF System to Random


Gaussian Force
The iterative process is demonstrated for the beam that was represented
as an SDOF in Fig. 6.2. For this beam it was found that

Therefore,
Purchased from American Institute of Aeronautics and Astronautics

88 G.MAYMON

According to Eq. (7.10)

defining
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where ull is the standard deviation of the response.


The equivalent rigidity single term matrix is degenerated into

[k + k,,] = kIl + k;? = 38.8873 + 1160.9172~:~ (7.20)

For an SDOF system, the normal mode is always equal to one

and the generalized mass

Assuming that the excitation force F has a one-sided white noise spectral
density of value S1 = 0.065 kg2/(rad/s), the matrix S4(w) is, therefore,

and the spectral density of the response is

The variable lleq in Eq. (7.24) is the equivalent modal damping coefficient.
For the original system, the damping is assumed to have no nonlinear
effects. As a result, cll in Eq. (7.16) is a constant, although this is not
necessarily true for the modal damping coefficient. This coefficient is defined
for a linear SDOF system as

and for the present case


Purchased from American Institute of Aeronautics and Astronautics

STATISTICAL LINEARIZATION METHOD 89

Therefore, during the iterative solution process, lie, changes as ole,


changes, so that

where is the damping coefficient of the linear system and llIin = 0.02
for this case (see Fig. 6.2). The mean square value (the variance) of the
response is given by Eqs. (7.14) and (7.24). Denoting this (T: (a, being the
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

standard deviation of the response) the following i s obtained:

Using Eq. (7.26) this yields

The resonance frequency is calculated by the characteristic equation

or, by introducing the numerical values,

Therefore, the iterative process is as follows:


1) Calculate allinfor the linear system

olli, = d38.887319.4118 X = 203.2668 radls

2) Calculate (T: by Eq. (7.27b)

3) Substitute this value into Eq. (7.28) and calculate a new value for
the frequency
Purchased from American Institute of Aeronautics and Astronautics

90 G. MAYMON

4) Calculate new CT; with Eq. (7.27b)

5) Repeat the process until convergence is obtained.


In Fig. 7.1 the consecutive values of cri and ole,are shown. The values
are also repeated in Table 7.1.
Because the response is sought, convergence criteria should be applied
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

to the mean square of the response but other parameters such as the
equivalent frequency and the equivalent rigidity should also be checked.

Iteration

0 5 10 15 20 25 30 35 40 45 50
Iteration
Fig. 7.1 Convergence of the variance and the equivalent frequency.
Purchased from American Institute of Aeronautics and Astronautics

STATISTICAL LINEARIZATION METHOD

Table 7.1 Iterative converging process, SDOF

Iteration kfP Weq (4


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The solution converges to

i.e., the standard deviation of the Gaussian response of the mass (or the
midspan of the beam) is

with a governing frequency of 393 radls. It should be noted that the standard
deviation of the linear system is
a, = 0.5858 cm
Purchased from American Institute of Aeronautics and Astronautics

92 G.MAYMON

with a governing frequency of 203 radls. Thus, the nonlinear response is


about 52% of that calculated by a linear analysis, emphasizing the fact that
a less conservative result is obtained using a nonlinear analysis. The shift
in the frequency (hardening) is also significant.
A closed-form analytical solution for an SDOF oscillator is presented in
Ref. 24. For the present case the solution yields the following result for
the variance of the response:
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

which differs from the numerical iterative result by only 1.2%.

111. Nonlinear Random Response of a Two-DOF System to Random


Gaussian Force
A nonlinear two-DOF system is described in Fig. 7.2. Each spring is
nonlinear, with its elastic restoring force equal to

The viscous dampers are linear. The damping is cl = 2101% and


c2 = 2 l o 2 S , where lol and co2
are the damping coefficients of each
damper (and not the modal damping). Also yl and y, are relative to the
fixed base. It can easily be shown that the equation of motion of this system is

Fig. 7.2 Two-DOF system.


Purchased from American Institute of Aeronautics and Astronautics

STATISTICAL LINEARIZATION METHOD 93

It is more reasonable to select general coordinates {q) so that

where ql is the displacement of the first mass relative to the fixed base and
qz is the displacement of the second mass relative to the first mass. It
should be noted that elastic forces in the springs (and, therefore, stresses if
required) can be more easily calculated with these generalized coordinates.
Equation (7.30) then takes the form
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

These equations can take a simpler form in which the damping matrix is
symmetric, if both sides are multiplied by a matrix

which yields

These are the equations for which equivalent linearization is performed.


The nonlinear functions @ are

and, therefore, according to Eq. (7.10)

= 3clklE(q:) =

= 3c2k2E(qz) = 3c2k2ffz (7.35)

where CT: and a$ are the variances of ql and q2, respectively.


Purchased from American Institute of Aeronautics and Astronautics

94 G.MAYMON

The normal modes matrix of this system is


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The generalized matrix is

The generalized damping matrix is

and the modal damping coefficients are

where CI1 and CZ2are the diagonal terms of [q, M1 and M2 are the
generalized masses, and wl and w2 are the natural frequencies.
Assuming that forces Fl and F2 are uncorrelated zero mean stationary
Gaussian processes, with one-sided white noise spectral densities of magni-
tudes S1and S2,the spectral density matrix of the excitations for Eq. (7.33) is

The generalized spectral density matrix S+(w) is given by Eq. (7.12), and
the spectral densities of the response at the first and second masses are

where [&I, [+2] are given by Eqs. (7.36).


To demonstrate the solution process, numerical values are introduced:
ml = 0.010 kg . s2/cm (weight of 9.80 kg), rn2 = 0.015 kg . s2/cm (weight
Purchased from American Institute of Aeronautics and Astronautics

STATISTICAL LINEARIZATION METHOD 95

of 14.7 kg), k1 = 1000 kglcm, k2 = 500 kglcm, q k l = 200 kg/cm3, e2k2=


100 kg/cm3, lol= lO2 = 0.01, and S1 = S2 = 180 kg2/(rad/s).
The characteristic equation of the undamped linear part of Eq. (7.33) is
solved for the linear natural frequencies and the normal modes
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

which yields the following results:

= 143.06 radls;
qlin wzlin = 403.567 radls

By Eq. (7.37)

Using Eqs. (7.38) and (7.39), the linear modal damping coefficients are ob-
tained

Substituting the normal modes and the input spectral density (7.40) into
Eq. (7.12) yields
Purchased from American Institute of Aeronautics and Astronautics

96 G. MAYMON

Using the procedure described in Chapter 5, Sec. 11, the following expres-
sions are obtained for the mean square values (variances) of ql and 9,:
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

In Eq. (7.47) H(o) is equivalent to H(w) as M I and M2 were already


introduced in Eq. (7.46). The values of the integrals can be calculated by
the expressions given in Appendix A.
The results obtained for the linear system are

It should be noted that products like leqo;,and similar are presented in


the denominator of the frequency integrals. The calculation for systems
with linear damping can be shortened if the term ~ l i n ~ l i , is~ ~introduced
,
instead of this product [see Eq. (7.28)]. Therefore, the values obtained in
Eqs. (7.43) and (7.45) can be used during the iterative process, saving
repeated computations of l,,.
The iterative process was implemented on a personal computer, using
the TK+ Mathematical Solver program. Because only two DOFs exist, the
characteristic equation was solved explicitly as a quadratic equation and
then the mode shapes were calculated. Multiplication of matrices was done
explicitly. When more DOFs exist, an eigenvalue and eigenvector subrou-
tines can be used, as well as matrix multiplication modules. Other mathemat-
ical solver programs can also be used.
In Table 7.2, results of the first 20 iterations are presented for o l , M I ,
&, and 4and 02, M2, h 2 , and (T:. Here +11 and 421are defined as unity.
The hardening effect can be observed clearly by comparing the final
nonlinear frequencies

to the corresponding linear frequencies


Purchased from American Institute of Aeronautics and Astronautics

Table 7.2 Iterative converging process, 2 DOFs


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Iteration 01 412 MI d 0 2 422 M 2 d


Purchased from American Institute of Aeronautics and Astronautics

98 G. MAYMON

Also, the rms (standard deviations) of the responses of the nonlinear sys-
tem are

al = = 1.6981 cm; a2= = 2.0797 cm

These are significantly lower than those of the linear system

= 2.5643 cm; a2= = 3.6132 cm


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

crl =

IV. Nonlinear Random Response of an Elastic System


As the procedure described in the preceding section is based on finite
element solution of the coefficients of nonlinearity, and normal modes are
used in this procedure, there is no loss of generality in describing a beam
structure as a representative example of other continuous elastic
structures.
In Chapter 6, Sec. 111, the linear rigidity matrix [ki,][Eq. (6.12)] and the
nonlinear rigidity matrix [bij][Eq. (6.14)] of a simply supported beam were
calculated, using a finite element program. Nonlinear response of this beam
is now calculated, assuming the beam is subjected to a midspan random
exciting force with one-sided white noise spectral density of value S2 = 0.1
kg2/(rad/s). Linear modal damping coefficients of the three first modes are
assumed as = l2= l3= 0.02. The beam is represented by a three-DOF
equivalent system (shown in Fig. 7.3), where the total mass of the beam
MT was lumped as shown. Other methods of lumping masses are also
possible. The treatment of a beam is used to demonstrate the process of
a nonlinear solution of an elastic structure.
Using Eqs. (7.10), (6.12), and (6.14), the equivalent linearized rigidity
matrix is

The expected mean square values a $ l , and CT:~,at ml, m2, and m3,
respectively, are to be calculated. Because of beam symmetry, a;, = CT:~,
only two mean square values, a:, and a:2, are calculated. The cross-spectral
response between y, and y2 is also calculated.
Purchased from American Institute of Aeronautics and Astronautics

STATISTICAL LINEARIZATION METHOD


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 7.3 Three-DOF equivalent beam.

The characteristic equation from which the equivalent natural frequen-


cies and normal modes are calculated is

This equation is solved to yield three natural frequencies, wl, w2, and w3,
and three normal modes, with the following matrices:

It is convenient to define the normal modes so that

The generalized masses are obtained by


Purchased from American Institute of Aeronautics and Astronautics

100 G.MAYMON

The spectral density matrix of the excitation is

[SQ(~)]= [:z0 0 0
(7.52)
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The procedure, which was already applied for two DOFs is repeated for
three DOFs. The generalized spectral density matrix is

The matrix B = [H(o)] [S4(o)][H*(w)] [see Eq. (7.13)] is

Here Hi is equivalent to Hi, because division to MiMj were already per-


formed in Eq. (7.53). The spectral density of the response at ml is

1 + sh21H212(421>2
sql(u) = s411H11~(411)~ + s433/H312(43i)2

+ 2S412421411Re(HTH2) + 2SdI3431411 Re(H1*Hd

+ 2S423d'21431 Re(H2*H3)
Purchased from American Institute of Aeronautics and Astronautics

STATISTICAL LINEARIZATION METHOD 101

and the mean square of the response is given by

The integral can be calculated using the expressions in Appendix A.


Similarly, the spectral density of the response at m2 is

+ fSaIH312($32)2
+ S4221H212(422)2
Sq2(w)= S4111H112(412)2
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

and the mean square of the response at this point is

The cross-spectral density between points 1 and 2 can be calculated using

Sq12= 1 4 1 1 [ ~ 1 1 4 2 l ~ (7.59)

where B is taken from Eq. (7.54). This yields

Sq12(0) = S411412411IHI I 2 + S4224224211H212+ S433


$32431 IH3 l2

and the expected value of yly2 is

The iterative procedure of solution is used again. The first step is to compute
the relevant quantities of the linear system. The first three natural frequen-
cies'' are

= 206.013 radls;
ollln = 807.132 radls;
ozlin = 1713.38 radls
031,n
Purchased from American Institute of Aeronautics and Astronautics

102 G. MAYMON

and the matrix of the normal modes is


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The variances and the covariance of the linear response are, therefore,

The iterative process is solved on a personal computer using the TK+


Mathematical Solver program. Eigenvalues (natural frequencies) were
solved using an explicit solution of a third power equation. If more than
three DOFs are taken into account, a different algorithm for the eigenvalues
and eigenvectors must be used. Other mathematical solvers may be used.
Results are described in Fig. 7.4. The hardening effect is clearly seen by
computing the linear and nonlinear frequencies and responses, as described
in Tables 7.3 and 7.4, respectively.
It can be seen that nonlinear analysis yields a response with rms values
(standard deviation) that are less than 50% of the calculated linear analysis
values. This justifies the claim that the use of nonlinear analysis is less con-
servative.

Table 7.3 Linear and nonlinear frequencies

Frequencies, radls

Mode 1 2 3

w, = linear 206.013 807.132 1713.38


a,,= nonlinear 446.456 930.289 1790.10
wndwi 2.1671 1.1526 1.0448
Purchased from American Institute of Aeronautics and Astronautics

STATISTICAL LINEARIZATION METHOD 103


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Iteration

Iteration
Fig. 7.4 Convergence a) of the frequencies and b) of the variances.

A designer should be aware of the importance of the rms of the response.


The response of a system subjected to a stationary Gaussian excitation with
zero mean is a stationary Gaussian process, with a mean of zero. Its stan-
dard deviation is the rms value. This implies that the response values are
less than the rms value in 66.23% of time (la), less than twice the rms
value for 95.45% of time (2a), and less than three times the rms value for
99.73% of time (3a). Thus, for the described example, the center of the
beam (point 2) can vibrate with responses that are lower or equal to
3 X 0.3214 = 0.9642 cm for 99.73% of the time. The distribution of the
maximum values of the response is treated more extensively in Chapter 12.
Purchased from American Institute of Aeronautics and Astronautics

1 04 G.MAYMON

Table 7.4 Linear and nonlinear rms values of the responses

Mean squares, cm2 Root mean squares, cm

a:? a:,? ~ Y I

Linear response 0.24568 0.49324 0.34690 0.4957 0.7023 0.5890


Nonlinear response 0.05262 0.10330 0.07263 0.2294 0.3214 0.2695
Ratio nonlinearllinear
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

0.2142 0.2094 0.2094 0.4628 0.4576 0.4576

V. Computational Procedure
The iterative procedure described in the preceding chapters for the solu-
tion of the nonlinear behavior of a beam is based on approximating the
continuous structural system by a discrete MDOF equivalent system. In
doing so, the use of the finite element program is limited only to the steps
reauired for the determination of the discrete elastic linear and nonlinear
matrices of the structure and the relationship between these matrices and
the variances and covariances of the response. Once these relationships
are formulated, the iterative computational procedure does not include
further use of a finite element code. This implies that the user selects a
finite number of DOFs that participate in the analysis. The fewer DOFs
selected, the easier and shorter the computational procedure. An efficient
estimation of the required finite number of D O F ~ must be based on the
user's experience and on intelligent estimations based on the particular
structure analyzed.
The flow chart in Fig. 7.5 describes the basic steps for a nonlinear analysis
of a continuous structure, with cubic elastic nonlinearity.

VI. Calculation of Stress Response


In Chapter 3, Sec. IV, it was shown that stresses in a vibrating structure
can be solved using the stress modes approach. Once stress modes for
a particular required stress are determined (analytically, numerically, or
experimentally), the stress response of a nonlinear vibrating structure can
be calculated by the same procedures used to determine the displace-
ment response.
During the development of the basic expressions for the displacement
response, multiplication of certain matrices with L4,j and L4ilTwere per-
formed twice. First, the spectral density of the excitation was premultiplied
by L4il* and postmultiplied by L&l to obtain L s ~ ~This
. was done to obtain
a matrix of the spectral density of the generalized forces. Second, the matrix
[B] was premultiplied by L4i] and postmultiplied by L4ilT.This was done
to determine the spectral density of the displacement response at point i.
When calculating the stress response, only the second step is performed
with the stress modes instead of the normal modes. The first step, which
Purchased from American Institute of Aeronautics and Astronautics

STATISTICAL LINEARIZATION METHOD

- 1. Determine the number of


degrees-of-freedom to participate in the
analysis -
+
I 2. Calculate, analytically or numerically, the
linear elastic terms of the discrete equivalent
system
I
1
-3. Calculate, analyticallyi or numerically, the
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

non-linear elastic terms of the discrete


equivalent system
4. If non-linear damping

*
damping on velocity
andlor dis lacement y
5. Formulate k? , c y , m y as function of
D
1 0;by the given equations I
I 6. Formulate solutions for Oi , [ ] and [MI
6a. Formulate solution for the stress modes, if I
7. Defme the external
power spectral density

1 8. Formulate expressions for the matrix S*;; I

11. Calculate O iand [ 1. First iteration- linear


system
1la. Calculate the stress mode, if required
1
I

12. Calculate SOijby the expressions of step 8


t
I
I next iteration 1
113. Calculate 0;withthe equations of step 9 (
I
14. Calculate new matrices for [k] [c] and [m]
1 no

Fig. 7.5 Computational procedure.


Purchased from American Institute of Aeronautics and Astronautics

106 G. MAYMON

determines the generalized spectrum of excitation, should be done with


the normal modes of the system, and not the stress modes. Therefore,
when applying the iterative procedure for a nonlinear stress response, both
normal modes and stress modes must be recomputed in each iteration.

VII. Summary
The statistical linearization method is presented. This method is based
on the solution of a set of linear differential equations obtained by minimiz-
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

ing the difference between the original nonlinear equations and the equiva-
lent equations.
A Gaussian random excitation with zero mean is assumed.
The statistical linearization method yields an iterative procedure for the
computation of the response mean square values. This iterative procedure
can be automated, and was demonstrated for SDOF, MDOF, and elastic
systems.
The demonstrated results confirm the conclusion that nonlinear analysis
may result in a less conservative design.
Purchased from American Institute of Aeronautics and Astronautics

Chapter 8

Nondeterministic Structures: Basic Concepts


I. Introduction
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

T HE design of any engineering system is a process of decision making,


under constraints of uncertainty. The uncertainty in the design process
is due to the lack of deterministic knowledge of different physical parame-
ters and the uncertainty in the models with which the design is performed.
This is true for all disciplines such as electronics, mechanics, aerodynamics,
and structures involved in any design.
The uncertainty approach to the design of subsystems and complete
systems was advanced by the concept of reliability. Systems are analyzed
for possible failure processes and criteria, probability of occurrence, reliabil-
ity of components, redundancy, possibilities of human errors in the produc-
tion, and additional uncertainties. According to this type of approach, a
total required reliability of a certain design is defined with proper reliability
appropriation for subsytems. This required reliability certainly influences
both the design cost and the product cost.
Nevertheless, in most cases the structural analyst is still required to supply
a design with absolute reliability, and most structural designs are performed
using deterministic solutions. To compensate for the lack of knowledge,
structural designers use a safety factor (lack of knowledge factor?), thus
recognizing de facto the random character of many design parameters.
During the last decade, the need for application of probabilistic methods
for nondeterministic structures started to gain acceptance within the struc-
tural design community. Designers started to adopt the stochastic approach
and the concepts of structural reliability. It is likely that by the end of the
decade, this approach will dominate structural analysis procedures and,
thus, the structural design will become more integrated in the total sys-
tem design.
The main sources of uncertainties in structural analysis and design are
the model, the loads, and the uncertainties in various structural parameters.
The first two were briefly discussed in the Preface.
The treatment of uncertainties in structural parameters (stochastic struc-
tures) has become extensive and practical only in the last 10-20 years,
although pioneering studies were published earlier, in the late 1960s and
the early 1970s, especially by scientific journals of the civil engineering
community. Tremendous progress has been made in the formulation of
mathematical models and the establishment of several algorithms for the
determination of the behavior of a stochastic structure submitted to excita-
tion of stochastic loads. This progress was followed by the adoption of the
relatively new technology for practical application in the structural design
Purchased from American Institute of Aeronautics and Astronautics

108 G. MAYMON

process in industry. It is likely that in the next decade more and more
design codes and specifications will include the use of probabilistic analysis.
The introduction of these methods into practical applications will be quicker
than the introduction of finite element programs, because of the large
infrastructure that exists in computational structural analysis and the rapid
advances in computational power.
For many years nondeterministic structures were tested using the Monte
Carlo simulation. A problem is solved many times, each time with a set of
deterministic values of the system parameters. These values are selected
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

randomly from the legitimate space of the basic variables, including the
approximate knowledge of the probability distribution functions of these
variables. In some simulations, success of the system is detected, and in
other simulations the system fails. The ratio between the number of failures
to the total number of simulations is used to estimate the probability of
failure of the system. The main disadvantage of the Monte Carlo simulation
is that large numbers of deterministic simulations are required, especially
in the tails of the distributions. As very low probabilities of failures are
required in structural analysis, lo4-106 simulations are required for a typical
problem. If one simulation run of a complex structure (using a finite element
program, for example) is in the order of several minutes to hours, the use
of the method becomes prohibitive for practical industrial applications as
a result of the large computation time that interferes with the schedule of
any project.
In the last decade, numerical algorithms that use nonsimulative methods
were developed. These resulted in several computer programs that solve
the probabilistic structural analysis problem within a reasonable and practi-
cal time frame and that are suitable for industrial use. Some of these
programs are briefly described in Chapter 9, Sec. IV.
Some of the numerical examples, which are introduced later in this book,
provide the reader with the basic concepts of probabilistic structural analysis
and the methods that can be applied for solution of practical structural
analysis problems. As most of the theoretical methods are covered exten-
sively in several textbooks, e.g., Refs. 11 and 25-27, many conference
proceedings and scientific papers, e.g., Refs. 28 and 29, only very basic
mathematical evaluations are repeated in this text.
The following simple example is introduced to clarify the difference
between the traditional factor of safety approach and the concept of proba-
bilistic analysis.
Assume that an elastic bar of rectangular cross section A = b X h is
subjected to a tensile force F, and assume that b, h, and F are normally
distributed, with the following standard deviations:

where a() is the standard deviation of ( ) and ,u( ) is its mean.


Purchased from American Institute of Aeronautics and Astronautics

NONDETERMINISTIC STRUCTURES: BASIC CONCEPTS 109

The stress in the bar is given by

and is approximately normally distributed, with a mean value psb=


2083 kg/cm2 (arbitrarily selected) and a standard deviation as =
104 kg/cm2. The maximum tensile stress in 3 a (three standard deviations) is
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

and in four standard deviations is

The bar is assumed to fail if the tensile stress in it is equal to or higher


than the yield stress. Assume that the yield stress Sy of the bar material is
normally distributed, with the following mean and standard deviation:

Thus, the minimal values of S, for three and four standard deviations are

Three types of designers who would treat this problem differently are iden-
tified.
1) The nominal designer calculates the safety factor (FS) of the bar from
the nominal mean values

If this value is equal to or higher than the required FS for the specific
project, the structure is approved. This approach does not take into account
the dispersion expected during the production phase in the cross section
dimensions b and h, the possible dispersion in the material property S,
and the uncertainty in the external force F.
2) The 3a worst-case designer calculates the safety factor by
Purchased from American Institute of Aeronautics and Astronautics

110 G.MAYMON

which is not acceptable. Two kinds of design corrections are possible: a)


Increase the cross section by 26% (1.210.95 = 1.26). This increases the
weight by 26% and, subsequently, the product cost. b) Decrease the allowed
tolerances on b and h (the designer cannot control the dispersion in F and
S,). A simple calculation shows that to obtain FS = 1 (not the required
1.2), tolerances on the cross section should be decreased from 1%to 0.04%.
This increases the product cost tremendously. A combination of a and b
is also possible.
A 4c~ worst-case designer faces much more difficult problems.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

3) Assuming that the required reliability of this specific project is 99.9%,


the probabilistic designer uses the probabilistic procedures described later
to show that the probability of the stress in the bar to be equal to or lower
than the yield stress is

Pr(Sb I S,) = 99.943% (8.6)


This value is higher than required, and the designer is allowed to slightly
decrease the cross section or slightly increase the required tolerances.
Although this example is a simplification of real-life applications, it shows
that contradictory measures can be taken when using each approach. It
also indicates that the probabilistic approach may be more realistic and
may result in less conservative designs. This approach is valid if project
requirements are formulated by probabilistic methods. Although this is not
the situation at present, it is believed that future requirements will contain
more and more of this approach.

11. Failure Surface: Basic Case


A failure surface g(R, S) is defined as a function of all random variables
of the problem, so that on one side of this surface the structure is safe and
on the other side it fails. The basic case includes two random variables. R
is the allowable quantity (stress, displacement, etc.) in the structure and
is designated the resistance term. S is the actual quantity in the structure
and is designated the load effect term. Traditionally, a negative value of
g(R, S) means failure and a positive value provides a safe region. The
failure surface is then

This is a linear expression and for the basic case, the failure surface is a
straight line, as shown in Fig. 8.1.
A safety margin can be defined as

so that if M 5 0, the structure fails and if M > 0, the structure is safe. The
failure surface is then
Purchased from American Institute of Aeronautics and Astronautics

NONDETERMINISTIC STRUCTURES: BASIC CONCEPTS 111


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 8.1 Failure surface of the basic case.

Assume that R and S are normal random variables with E ( R ) and E(S)
mean values and D ( R ) and D ( S ) standard deviations. Therefore, M is
also random, with E ( M ) as its mean and D ( M ) as its standard deviation.
Schematically, the failure and safe regions can be described on one-dimen-
sional axis, as shown in Fig. 8.2.
E ( M ) can be described in terms of the standard deviation D ( M ) ; thus,

where 0, is called the reliability index. The subscript c is after C0rnell,3~


who first formulated the problem in this form. Higher values of 0, describe
systems with higher reliability, or systems with smaller probability of failure.
Assume that R and S are normally distributed. Then M will also be
normally distributed, with the probability density function &M), shown in
Fig. 8.3. The shaded area is the probability of failure, the probability that
M will be smaller than or equal to zero.
The probability of failure is defined as
Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 8.2 Fail and safe regions, basic case.

which can be written as

The first term in the brackets is the transformation of M into a standard


normal distribution (mean equals zero, standard deviation equals one),
which is denoted u

The second term in the brackets equals -6, [Eq. (3.4)]. Therefore,

where cP is the standard normal CDF. Note that the original failure functions
(8.7) and (8.8) are linear. The result (8.13) is obtained because of this
linearity. Thus, the probability of failure of the system can be calculated
once the reliability index is known.
Assume R is normally distributed having PDF &(r) with E(R) and D(R)
as mean and standard deviations, respectively, and S is normally distributed
Purchased from American Institute of Aeronautics and Astronautics

NONDETERMINISTIC STRUCTURES: BASIC CONCEPTS 113


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 8.3 Normal PDF of M.

&(s) with E ( S ) and D(S). The joint PDF is &S(r, s). The probability of
failure is the volume of this function over the failure region, as shown in
Fig. 8.4. Thus,

Assume that R and S are independent variables, then

Thus,

@R(s) . &(s) ds (8.16)

where DR is the CDF of r.


Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 8.4 Basic case, two normally distributed RVs.

The integral (8.16) can only be solved explicitly in very few cases. For
the two-variables basic cases presented, it can be shown that Pf can be
calculated using Eq. (8.13), where

The random variables R, and S can be transformed into standard normal


variables U R , and us, respectively, using

Thus,
Purchased from American Institute of Aeronautics and Astronautics

NONDETERMINISTIC STRUCTURES: BASIC CONCEPTS 115

and the failure function g(R, S) can be transformed into another function
G(uR, us) to yield

This function is described in Fig. 8.5.


It can be easily shown (for instance, by trigonometric relations of lines
in Fig. 8.5) that the reliability index P is equal to the minimum distance
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

between the failure surface and the origin in the transformed variables
system. The point on G that has minimum distance to the origin is called
the design point, or the point of most probable failure, and is denoted by
asterisks, u i and u:. The term design point, which is used extensively in
the relevant literature, is misleading; it is somewhat incongurous to call the
point of most probable failure the design point.
The preceding analysis can be extended to n random variables. The
failure surface is then a function of X, variables

and the transformed failure surface is

Both functions describe a hypersurface of n dimensions. The exact probabil-


ity of failure is obtained by extending Eq. (8.16) to

where ,,..., x, (xl, x2,... ,x,) is the joint PDF of n random variables.
The reliability index is the smallest distance from the origin of the
transformed (standard normal) space to the transformed hypersurface
G(ul, 242, . . . , u,) = 0.
If the failure surface is linear, it forms a hyperplane of n dimensions,
both in the original and in the transformed spaces. If the failure function
is nonlinear, a hypersurface is obtained and P is the distance to a hyperplane
tangent to this hypersurface at the design point, the point closest to the
origin of the transformed space.
It has already been shown that instead of performing the integral (8.22)
of the joint PDF over the failure region G 5 0, the reliability index P is
calculated by determining the closest distance of the transformed surface
to the origin, and by using Eq. (9.13) to calculate the probability of failure.
This is based on the replacement of the failure hyperspace with a tangent
hyperplane. Thus, Eq. (8.13) is equivalent to performing the integration of
the joint PDF over the space marked first-order reliability method (FORM)
in Fig. 8.6. If the failure surface is nonlinear, the integration should be
Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

SAFE

FAIL

Fig. 8.5 Transformation into standard normal space, a) physical space and
b) standard normal space.

done over the space EXACT in the figure. In second-order reliability


methods (SORM), the hyperspace is replaced by a quadratic hypersurface,
and the integration should be done over the space SORM in the figure.
The difference between FORM and SORM results depends on the non-
linearity of the failure surface. In most practical cases, in structural analysis,
very low probability of failure should be allowed; thus, the design point D
is in the tail of the joint PDF. The main contribution to Pf,therefore, is
Purchased from American Institute of Aeronautics and Astronautics

NONDETERMINISTIC STRUCTURES: BASIC CONCEPTS 117


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

' 'second Order Design Point


Curve

Limit Function
Fig. 8.6 Exact, FORM, and SORM integration spaces.

from the area above the tangent line and, thus, FORM methods give usually
good approximations of the probability of failure. Expressions relating Pf
and 0 for SORM were also developed and presented in the literature.
These expressions contain radii of curvature of the failure function.

111. Reliability Index


In 1969, Cornel13' suggested the following formulation for the reliabil-
ity index:

where pz is the value of the failure function g(Xl, X2,. . . , X,) at the mean
values of all of the random variables; thus,

and
Purchased from American Institute of Aeronautics and Astronautics

118 G.MAYMON

where the derivative of g with respect to the variables Xi are taken at the
mean point pi.These definitions are extensions of the basic definitions in
Eq. (8.10).
A major disadvantage of the Cornell reliability index is that it is not
invariant to the selection of the failure surface, when several possibilities
for formulations of this function are possible for the same failure criterion.
This was demonstrated in Ref. 25 for the following example.
Assume a bar of cross section A on which a tensile force F is acting.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Assume that the allowable stress in the bar is R. Two failure functions can
then be formulated.

Suppose that R has a mean pR = 62 and standard deviation vR = 6.2, and


A has a mean pA = 2.8 and standard devation a, = 0.14. Also assume that
F = 100 is a deterministic force. In Table 8.1, Eqs. (8.23-8.25) are applied
to both gl and g,. It can be seen that the reliability index using these
different functions is not identical.
In 1974, Hasofer and Lind31 suggested a method to overcome the disad-
vantage presented by the Cornell reliability index. The random variables
Xi are transformed into standard normal variables ui by

The failure function g(X) = 0 is transformed into G(u) = 0, and the


smallest distance between this surface and the origin is found

pHL= minimum Jz u:; given G(ui) = 0

Both Cornell and Hasofer-Lind methods use the first and second moments
of the distribution of the random variables, but do not take into account
the real distribution of these variables. As the probability of failure of a
practical structure should be small, this happens in the tail of the distribution
functions. Thus, this method yields good results when the PDF of the
random variables are normal or close to normal at the tail.
In 1978, Rackwitz and Fiessler3' extended the Hasofer-Lind concept to
include, approximately, the PDF of the basic random variables. According
to this method, nonnormal variables are transformed into normal variables
in such a way that the PDF and the CDF of the original and transformed
variables are equal at the design point. As this point is not known in
advance, an iterative procedure is used to calculate the reliability index.
Purchased from American Institute of Aeronautics and Astronautics

NONDETERMINISTIC STRUCTURES: BASIC CONCEPTS 119

Table 8.1 Differential reliability index for different equivalent failure functions

Case 1 Case 2
- -

g g, = R - (FIA) gz = RA - F
PZ F PRPA- F
PR- - = 62 . 2.8 - 100 = 73.60
PA
100
= 62 - -= 26.286
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

2.8
cz .\/a; + ( F l d ) ' . u, = 6.452 dpic i + pi
0;= 19.41

P C 26'286 - 4.07
-- --
73.6 - 3.79
6.452 19.41

The algorithm which is used in the Rackwitz-Fiessler method is as


follows:
1) Define the basic physical random variables Xi(i = 1, 2, . .. , n), the
cumulative distribution F, and the PDF f;: of each variable. Each variable
has a mean piand a standard deviation ai. Define the failure surface g(Xi)
= 0 of the problem.
2) Transform each variable Xito a standard normal variable ui using

and define the transformed failure function

3) Find the minimum distance between the surface g, and the origin

/3 = minimum d u : + ut + . . . + uz; given gl = 0 (8.31)

This solution defines the first design point u?,u;, . .. ,u,*.


4) Calculate XF = uTui + pi.
5 ) Calculate the mean value and the standard deviation of an equivalent
normal distribution of each basic variable that is not normal, using
Purchased from American Institute of Aeronautics and Astronautics

120 G. MAYMON

where @ is the standard normal CDF

@(z) = 0.5 + erf(z)

and 4 is the standard normal PDF

1
(-i
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

+(z) = -exp z')


6
6) Define new standard normal variables

(8.35)

7) Calculate the reliability index Dl as

pl = minimum VET+ E$ + . . . + E;; given g1(E?) = 0 (8.36)

-
8) Repeat steps 5-7 until a convergence criterion, such as 5
E, is obtained. The last value /3 is the calculated reliability index. Experience
shows that convergence is obtained quite quickly. An example of this type
of iterative process is demonstrated in Chapter 9. A schematic description
of the method is in Fig. 8.7.
In 1981 Chen and Lind33 extended the Rackwitz-Fiessler method by
equating the slope of the PDF and the transformed standard normal PDF
at the design point. In many cases this method gives better results for the
reliability index.
In 1981 Hoenbichler and rack wit^^^ suggested the use of the Rosenblatt
transformation in the analysis of stochastic structures. According to this
method, each random variable can be transformed into a standard normal
variable. This is also done to dependent variables, which are transformed
into a system of independent standard normal variables. Therefore, a system
that has had many random variables of arbitrary PDFs and interdependence
of variables is transformed into standard normal independent variables.
The transformed system is solved by determining the design point as the
point closest to the origin in this independent variables space, and the basic
variables are then calculated by inverse transformation. This is the method
used today in most of the computational tools that exist for the solution
of stochastic structures.
When all of the basic variables Xiare mutually independent with marginal
CDFs Fxr(x,), the transformation is defined by
Purchased from American Institute of Aeronautics and Astronautics

NONDETERMINISTIC STRUCTURES: BASIC CONCEPTS 121

(1) Define all physical random


variables and their CDF, PDF, means
and standard deviations.

r-
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

(2) Transform each physical variable


X into u-space using (8.29), and
prepare a failure function in the

?
(4) Calculate physical "design point"

(5) Calculate new means and standard


(6),(7) With results (4) and ( S ) ,
do another iteration (equivalent

Fig. 8.7 Rackwitz-Fiessler iterative process.


Purchased from American Institute of Aeronautics and Astronautics

122 G. MAYMON

where @ is the standard distribution. This means that

and the inverse transformation


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

When the basic variables are not mutually independent, the first variable
is transformed by using

The second is transformed by

where FX2(x2jx1)is the distribution function of X2conditioned upon Xl= xl,

Expressions for CDFs can be found in the literature. Several examples can
be found in Ref. 25.

IV. Summary
The rationale behind the probabilistic treatment of structures is described
and discussed. It can be seen that different design decisions can be obtained
using the traditional safety factors and the probabilistic approaches.
The basic case of failure surface is formulated, defining resistance terms
and load effect terms in the failure function.
First-order and second-order reliability methods are described.
The reliability index, which is a measure of the structural reliability, is
defined. Three different definitions, as suggeted by different authors, are
described and discussed.
Purchased from American Institute of Aeronautics and Astronautics

Chapter 9

Calculation of the Probability of Failure


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Introduction

T HERE are several methods by which practical calculation of the proba-


bility of failure can be performed. The simplest method is an analytical
solution of the problem. An analytical solution is possible only when the
failure function can be described by a closed-form expression and is practical
only when a relatively small number of random variables are involved.
In such a case a set of algebraic andlor transcendental equations can be
derived and solved. Sometimes the solution process itself requires a numeri-
cal computer program, and thus it is a semianalytical solution. When
the PDF functions of the random variables are normal, the Hasofer-Lind
method 31 can be used. When nonnormal distributions exist, the Rackwitz-
Fiessler iterative method32 is a better alternative.
When the number of random variables is large (say, more than 5-6), the
analytical or semianalytical solutions tend to be cumbersome. This is also
the case when correlations between variables exist. In this case it is recom-
mended to use one of the commercially available special-purpose probabi-
listic computer programs, four of which are described later in this chapter.
When the failure function cannot be written explicitly, both analytical
and existing computer programs fail to provide a practical solution. Unfortu-
nately, these are the most common cases. Practically the load effects term
S in the failure function (8.7) cannot be expressed explicitly for a practical
structure and is usually calculated using a finite element computer code.
The problem of incoporating the finite element codes and the probabilistic
programs will certainly be solved in the near future by interfacing these
two kinds of programs. Meanwhile, approximate explicit failure functions
can be derived using, for example, the Taylor series expansion method.
FORM solutions can be derived using the modified joint PDF described
in Chapter 11.
The recommended methods of solution are described in Fig. 9.1. Also
chapter numbers, which describe each method shown in the figure, are listed.
All of the commercially available probabilistic codes provide the possibil-
ity to calculate not only the probability of failure of a component, which
includes several random variables, but also probabilities related to combina-
tions of components into systems, where components in series and in parallel
exist. This subject is not evaluated here, and the interested reader can find
more details in related references, e.g., Refs. 25-27.
Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON

I
small number of
yes
I
Closed-form failure function

i
t
large number
+ I

large number of
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

variables, of variables, variables, including


no correlations including correlations
correlations

distributions distributions distribution


El distribution

i
Commercially Taylor series Joint
Multiplier sler Iterations available expansion to Modified
Method computer provide approx. Probability
programs closed-form Density
expression Function

I Method
t

solution solution SORM & solution solution


Simulations

Fig. 9.1 Methods of solution for the probability of failure.

11. Lagrange Multiplier Method


When the failure function is given by a closed-form expression, it is
transformed into a function of standard normal and mutually independent
variables. The minimum distance from the origin to the transformed failure
surface can be calculated using the Lagrange multiplier procedure.
The failure surface is
Purchased from American Institute of Aeronautics and Astronautics

CALCULATION OF THE PROBABILITY OF FAILURE

the distance between a point u l , u2, . . . , u, to the origin is

The following expression is constructed:

D = P2 - A . g ( u l , u 2,..., u,)
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where A is a Lagrange multiplier.


Equation (9.1) plus the following ( n ) equations

provide n + 1 equations for n values of u: plus the value of A.


As g(ui) is not necessarily linear, the n + 1 algebraic equations for the
n + 1 unknowns are also not necessarily linear. A general closed-form
solution is not possible but many numerical tools can be used to solve this
set of equations. Once the values of u* are known, the minimum value of
@ is obtained by Eq. (9.2). Also the values of the basic random variables
at the design point can be obtained by inverse transformation of uT.
Some simple examples demonstrate the use of the Lagrange multiplier
method for determination of the design point and the reliability index.

Example 9.1
A bar of random cross section A and a random yield stress S, is subjected
to a random tensile force F. Mean values, standard deviations, and PDFs
of these parameters are tabulated in Table 9.1.
The failure surface is given by

Transforming the basic variables into standard normal variables using

Table 9.1 Data for Example 9.1

Parameter Units Distribution Mean, p Standard deviation, u

SY kg/cm2 Normal 2500 75


F kg Normal 4166 125
A cm2 Normal 2 0.04
Purchased from American Institute of Aeronautics and Astronautics

126 G. MAYMON

yields the failure function in the u space

Equation (9.3) takes the form


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Equation (9.4) yields

Equations (9.9) and (9.7) are solved for u* to yield

Using Eq. (9.2) yields

The results in Eq. (9.10) indicate that the most probable point of failure
is the combination when the yield stress is 2.636963 standard deviations
less than the mean, the acting force is 2.273637 standard deviations more
than the mean, and the cross-sectional area is 1.675018 standard deviations
less than the mean. Using Eq. (9.6) for inverse transformation, the physical
quantities at the design point are
Purchased from American Institute of Aeronautics and Astronautics

CALCULATION OF THE PROBABILITY OF FAILURE 127


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

0.04 0.045 0.05 0.055 0.06 0.065 0.07 0.075 0.08


Standard Deviation of Area
Fig. 9.2 Reliability as a function of the standard deviation of the cross section.

Using first-order approximation, the probability of failure is

Thus, the reliability is

For parametrically varied standard deviation of the cross section, the relia-
bility of the structure is calculated by repeating the process just described.
The results are shown in Fig. 9.2. Thus, if a reliability of 99.9% is required,
a value of 0.078 cm2 for the standard deviation can be selected, a value
that is higher than the original design (Table 9.1), and the tolerances on
the cross-sectional area can be released.

Example 9.2
Assume that the acting force F of example 1 has a uniform probability
density between 3791 and 4541 kg. The first and the second moments
are, therefore,
Purchased from American Institute of Aeronautics and Astronautics

128 G. MAYMON

The cumulative distribution function of F is

F(F) = 0; elsewhere

Use of the transformation Eq. (8.37) yields


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

u2 = @-'[F(F)] (9.17)

that can also be written as

@ ( ~ 2 )= F(F)
0.5 + erf(u2) = 1/750[F- 37911; 3791 5 F 5 4541 (9.18)

Therefore,
F = 750[5.0546667 + erf(u2)] (9.19)

As S, and A in this example remain normally distributed with the same


data given in Table 9.1, the failure function in the u space is

and D is [from Eq. (9.3)]

The derivatives of D with respect to uiyield three equations

In the second equation of (9.22), the following relation was used


Purchased from American Institute of Aeronautics and Astronautics

CALCULATION OF THE PROBABILITY OF FAILURE 129

A solution of Eqs. (9.20) and (9.22) yields


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Pf(first order) = 0.04%

This reliability is smaller than that of example 1, due to the greater standard
deviation of the acting force and the greater probability to obtain higher
values of the force.

Example 9.3
Assume that the yield stress of the material in example 1 has a Weibull
distribution with mean and standard deviation equal to those given in
Table 9.1;

The Weibull distribution has the following CDF:

F(S,) = 1 - exp(-as$)

where a and j3 are constants, related to the first and second moments by

where r is the gamma function. Using Eq. (9.27) with the given moments
Eq. (9.25) yields

a = 8.18211 X j3 = 42.03862 (9.28)


Purchased from American Institute of Aeronautics and Astronautics

130 G.MAYMON

Using the transformation

~1 = @-'[F(Sy)]
@(ul) = 1 - exp(-as$)
0.5 + erf(ul) = 1 - exp(-as$)
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Therefore, the failure surface is

and

Differentiation with respect to ui yields

Solution of (9.30) and (9.32) yields

U: = -0.7743742 A* = 1.969025 crn

Pf(first order) = 0.075%


P, (first order) = 99.25%
Purchased from American Institute of Aeronautics and Astronautics

CALCULATION OF THE PROBABILITY OF FAILURE 131

111. Demonstration of the Iterative Process


In Chapter 8 Sec. I11 a solution process suggested by Rackwitz and F i e ~ s l e r ~ ~
is described. This process takes into account the true PDF of the nonnormal
random variables and provides an iterative process that may be simpler
for solution than the Lagrange multiplier method demonstrated in the
preceding section. In example 4 the iterative process is demonstrated.

Example 9.4
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Example 3 in the preceding section is solved using the Rackwitz-


F i e ~ s l e iterative
r~~ method. The steps taken are those described in Fig. 8.7.
1 ) The yield stress has a Weibull distribution given by Eqs. (9.25), (9.26),
and (9.28). The force F and the cross section A are identical to those given
in Table 9.1.
2 ) For the first iteration, the three physical variables Sy, F, and A are
transformed into ul, u2, and u3, respectively, as if they are all normal, using

The failure function g(ui) is, therefore,

3 ) Solving the algebraic equation obtained by using Eqs. (9.3), (9.4), and
(9.35) the following is obtained:

The value of fl is identical to that of example 1, where normal distributions


were assumed for all three variables.
4 ) Use of Eq. (9.34) yields

5) Using Eq. (8.32), new values are obtained for pll and ull, i.e., for the
first variable S,, which is the only nonnormal variable in this example,
Purchased from American Institute of Aeronautics and Astronautics

132 G.MAYMON

Table 9.2 Iterative converging process for Example 9.4

Interaction pl, ~1 l s,* F* A* P


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

6) These values are introduced into Eq. (9.34) instead of psy and gSy.
7 ) A new /3 is computed.
8) The iterative process continues until convergence is obtained. In Table
9.2, five iterations are listed.
The convergence criterion is determined as 16, - 5 0.001 and,
therefore.

This is equal to the results obtained for example 3 for the reliability index
and also for the three basic variables S,, F, and A at the design point.

IV. Numerical Programs for Probabilistic Structural Analysis


The numerical examples presented in the preceding section were solved
analytically for the first-order probability of failure. They were character-
ized by the small number of random variables and by the existence of a
closed-form expression for the failure function. This solution processes may
give the designer an insight into the effects of different random variables
and their distribution functions.
For a large number of random variables, the described method of solution
is cumbersome. In many practical cases, the randomness of some parameters
can be expressed only as random fields rather than random variables.
The physical parameters may also be interdependent so that the analytical
methods fail to provide a reasonably fast solution. In these cases more
sophisticated methods of solution may be required.
In many practical cases it is also important to calculate the sensitivity of
the reliability index and the probability of failure to small changes in
statistical moments of the random parameters. By calculating these sensitivi-
ties, the designer obtains a quantitative measure of the effect of mean
values (basic design) and standard deviations (tolerances) on the reliability
of the structure. This can focus the designer's attention on the important
variables that should be treated in the design process. Therefore, a complete
probabilistic analysis of a structure should focus not only on the reliability
index, but also on sensitivities and importance factors, an analysis that may
enable the designer to efficiently modify the design.
Purchased from American Institute of Aeronautics and Astronautics

CALCULATION OF T H E PROBABILITY O F FAILURE 133

Several methods and algorithms that were suggested for a comprehensive


analysis of stochastic structures can be found in the relevant literature, e.g.,
Refs. 25-27. Based on these algorithms, several computer programs were
developed; some were prepared in the academic institutions for local re-
search and others by research and development institutes and companies,
as a part of practical research efforts intended to develop design tools for
the engineering community. Some of the developed computer codes are
already commercially available, and a few will be described in brief in this
section. It should be emphasized that this section is not a complete survey
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

of available computer programs in this field.


The most complete commercially available program for the analysis of
probabilistic structures is PROBAN,35developed in Norway by Det Norske
Veritas, in cooperation with academic institutions in Norway and Denmark.
Version 3 of the program has been available since 1994.This is an interactive
program, suitable for many computer and workstation environments, used
for probabilistic structural analysis by means of FORM, SORM, and several
simulation methods, including Monte Carlo. Systems in series and in paral-
lel, as well as basic components, can be solved. The user must provide a
failure function by using many library functions as well as an option to
program a user's function and link it to the program. Thus, a compiler and
linker program should be available on the users network. Templates for
the user limit functions are included in the program. Sensitivities and impor-
tance factors are calculated. Output files and graphic displays can be printed
and plotted on standard peripheral equipment. A large variety of distribu-
tion functions is included in a library, and user-defined distributions can
also be treated. Random fields can be included after performing a proper
discretization into random variables. Correlations between various vari-
ables defined by the user are also treated and dynamic problems solved.
The main practical disadvantage of this program is its price, more than
$30,000. More information on PROBAN can be obtained by writing Det
Norske Veritas, Hovik, Norway.
Another commercially available program is the developed
in Canada by Martec, Ltd. The main features of this program are similar
to those of the PROBAN. The data should be prepared on an input file,
which takes considerable time with an inexperienced user. An interactive
data input is available, but it is not very simple to use. The developer also
has an easy interface with a commercial finite element program and has
stated that an easy interface to any other finite element program can be
provided. This program demands a specific compiler (Lahey Computer
Systems, Inc.). Templates for user limited function are provided. The cost
of the program is $3500-$5000, and it can be installed on a personal com-
puter. A Windows 95 compatible version is being prepared at this writing.
More information can be obtained from Martec, Ltd., Halifax, Nova Sco-
tia, Canada.
A program, which can be commercially purchased, the CALREL,37was
developed in the University of California at Berkeley, California. This
program is not interactive, and introduction of data is somewhat cumber-
some. The user has to program the failure function using Fortran language
Purchased from American Institute of Aeronautics and Astronautics

134 G. MAYMON

and, therefore, a Microsoft Fortran compiler and linker are required. No


graphic output is available in the 1993 version. The program serves as a
tool in the research programs at the University of California, and in this
context an interface between the program and a local finite element program
FEAP was reported. This allows the reliability analysis of structures for
which the load effects term has no closed-form expression and can be
treated only by a finite element program. The finite element program is
not one of the codes frequently used by the industry. The price of the
CALREL is approximately $1000. More information can be obtained from
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

the Earthquake Engineering Research Center, University of California,


Berkeley, California.
A large-scale development effort is being made in the United States in
the NESSUSIPSAM program. This effort is led by NASA Lewis Research
Center, in cooperation with the Southwest Research Institute in San Anto-
nio, Texas, and several universities. The program3' is reported to include
a finite element module as a part of the probabilistic program itself. This,
paradoxically, may be a disadvantage because most practical users in the
industry use one of the commercially available finite element codes such
as NASTRAN, ANSYS, and ADINA, and these cannot be changed easily.
The NESSUS program (and some derivatives that are suitable for personal
computers) is now available for lease for an annual fee by Southwest
Research Institute. More information can be obtained from Southwest
Research Institute, San Antonio, Texas.
There are some other probabilistic structural analysis programs reported
in the literature, most of which were developed in the academic institutions
by researchers who have a need for them in their research efforts. Main
researchers in this field work in Austria, Germany, Italy, and Japan.

V. Summary
Several methods of calculation of the probability of failure of structures
are described.
A Lagrange multiplier method that enables analytical calculation of the
first-order probability of failure is demonstrated with three examples.
The Rackwitz-Fiessler3* iteration method is demonstrated with one ex-
ample.
Several commercially available computer programs for probabilistic anal-
ysis of structures are presented and discussed.
Purchased from American Institute of Aeronautics and Astronautics

Chapter 10

Taylor Series Expansion of the Failure Surface


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Introduction

A S EXPLAINED in the preceding chapters, no closed-form expression


is available for the failure surface of most practical cases treated in
the realistic design process. Thus, the commercially available probabilistic
codes cannot be used directly. A finite element code can be interfaced with
a probabilistic program but this type of interface is usually not practical
for the user in the industry, where finite element commercial codes are
usually rented and treated as black boxes. Although the interfaces could
be developed by the finite element developers, the present market demand
for such interfacing does not seem to warrant the effort. With the growing
interest in probabilistic analysis of structures, it is likely that in the future
such efforts will be economically justified and will, therefore, be initiated.
Meanwhile, the structural analyst in the industry is faced with the lack of
practical tools when this type of analysis is required.
One of the methods available to solve the present problem is to use an
approximate closed-form expression for the failure surface, based on a
Taylor series expansion of the unknown failure f ~ n c t i o n . ~ The~ - ~ coeffi-
l
cients of this expansion can be be determined by using the commercially
available finite element programs for several deterministic cases.

11. Taylor Series Expansion

The load term S(Xl, X2, . . . , XI) of the I random variables can be
evaluated into a Taylor series around an evaluation point (EP) as follows:

I (-
+C ( x i - XEpr)2
ax; E ,
Purchased from American Institute of Aeronautics and Astronautics

136 G. MAYMON

where SR are terms of higher order. If the evaluation is done in a range


close to the EP, the terms (Xi - XEPi)are small and SR is of third and
higher order of this small difference and can be neglected.
Simplification of Eq. (10.1) yields
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

+ xx
I-I I

i=l j=i+l
(Xi - XEP,). (Xj - XEP,)
a ( ~ - i / , ) ( i + ~ ) '+ j

The total number of terms in Eq. (10.2) is k + 1, where k = Z(I 3)/2, +


and k + 1 = ( I + 2)(1 + 1)/2.
It was found4' in several numerical solutions that mixed terms can also
be neglected in many practical cases. For these cases, the number of terms
in Eq. (10.2) is (21 + 1).
Further simplification of Eq. (10.2) yields

where

Yn = (Xi - X E P , ) ~ ; n = I + l , I + 2 ,..., 21 i = 1 , 2,...,I

and a. is associated with the EP itself.


A total of m deterministic solutions can be found for S with a finite
element program, where for each case the XEPiis increased or decreased
by AXi. These solutions can be represented by the vector { S }
Purchased from American Institute of Aeronautics and Astronautics

TAYLOR SERIES EXPANSION OF THE FAILURE SURFACE 137

where So is the deterministic solution for the EP. In a matrix form, the m
solutions for {S}can be written as
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where the matrix [Y] has m rows and k + 1 columns, which are associated
with the quantities AXi around the EP. Thus, the first column of [Y] consists
of ones, and the first row (less the first term) are all zeros, as this row refers
to the EP itself. ~ x ~ l i c i t l ;[Y]
, is

The optimal set of coefficients for Eq. (10.5a) (based on the minimum least-
square error) is given by solving42

which is a set of linear equations for the coefficients {a}.

111. Selection of the Evaluation Point


In Eq. (10.1), the load effects term S was expanded into a Taylor series
around the EPs {Xi} = {XEP,}, which were selected arbitrarily. Expansion
around the design point (the most probable point of failure) is required to
obtain a good approximation of the load effects term. The design point is not
known in advance, and the solution of the required probabilistic problem is
needed to find this point. In an iterative process, the probabilistic problem
is solved with a set of Taylor series expansions obtained using Eq. (10.6),
Purchased from American Institute of Aeronautics and Astronautics

138 G.MAYMON

a new EP is selected, and the process is repeated. Completion of the process


results in a condition where the design point is very close to the selected
EP, or at least is within the selected range of the Taylor series expansion.
The best selected EP is the design point, which is not known in advance.
Therefore, the first selected EP should be chosen with care, applying engi-
neering considerations, some experience, and some intuition.
Usually the influence of most random variables on the load effects term
can be predicted. In a structure that can be analyzed only numerically, the
structure input parameters necessary for the numerical solution (such as
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

loads, local and overall dimensions, material moduli, and other structural
characteristics) form a list of variables on which the load effects term
depends. The designer must decide which of the parameters in this list can
be considered as random, according to the given case. Each of the random
variables is associated with a PDF, a mean value, and a standard deviation.
These can be taken from previous data or assumed.
Generally, the designer can predict the general trend of influence of a
given variable on the load effects term, for example, an increase of external
loads results in an increase in stresses and an increase in thickness decreases
the stresses and deflections. Thus, the basic random variables can be divided
into nominator type and denominator type. In cases where the influence
of a certain variable is not clear, it can be clarified by a few determinis-
tic calculations.
In the design point, nominator type variables take values above the mean;
thus, the EP for these variables should be the mean plus a certain small
difference. Denominator type variables take values smaller than the mean;
thus, the EP for them should be the mean minus a certain small difference.
In most practical cases, the standard deviation of the variable is small, and
a good practical choice for the difference is the standard deviation; thus,

XEP;= pi + Ui; nominator type variables

XEP~
= pi - ffi; denominator type variables (10.7)

The EP range should run from the mean value upward for nominator
type variables and from the mean value downward for denominator type
variables. A reasonable choice is

pi - 2ai = XEpi- ui 5 XEP,5 XEPi ui = pi; + denominator type


(10.8)

Thus, an evaluation range of u is obtained for each variable above or below


the EP.
It is convenient to select the rn deterministic solutions by which the
vector {S) and the matrix [Y] are generated, using the following method:
Purchased from American Institute of Aeronautics and Astronautics

TAYLOR SERIES EXPANSION OF THE FAILURE SURFACE 139

1) The first deterministic solution is at the EP itself. Thus, the terms of


the first column of [Y] are ones, and the k terms in the first row are zeros.
2) Each random variable is changed by +AX and by -AX, keeping all
other variables at the EP. In this case, many terms in [Y] are zeros. This
procedure is called a two-point evaluation range. The form of the matrix
[Y] obtained using this method is described in Eq. (10.5b).
Once the coefficients { a } are determined with Eq. (10.6), the accuracy
of the approximated S function relative to the exact function can be deter-
mined by using Eq. (10.3) and comparing the results to the vector {S}
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

obtained by the deterministic finite element solution. It has been suggested39


that the existence of an extremum of S as a function of Xi can be checked.
Thus X,, is defined by

-- 0
ax,
If X,, is within the evaluation range, the accuracy of the approximated S
(say, S1 in Fig. 10.1) is lower than that of another approximation (say, S 2 )
whose extremum is outside the evaluation range as the effect of a random
variable on the load effects is monotonous. This check can be performed
before probabilistic computations are made. The EP and evaluation range
can then be changed before these computations, saving analysis time.
Another procedure that yields more accurate results for the approxi-
mated S function is to evaluate the {a}coefficients using a large number
of deterministic solutions. Instead of evaluating each random variable in
two points, +AX and -AX, an evaluation can be made for four points,
- 2 M , -AX, +AX, and +2AX. Although this increases the number of

A :,
True

2 EP-xi 8 EP L
Xici- ' EP+xi
Fig. 10.1 S with and without extremum within evaluation range.
Purchased from American Institute of Aeronautics and Astronautics

140 G. MAYMON

required deterministic solutions to 41 + 1, it significantly improves the


quality of the approximated S function. This procedure is called four-point
evaluation method.
Once an approximate closed-form expression is obtained for the load
term, the failure function can be approximated by the closed-form ex-
pression
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where X F is the resistance R random variable and the two last terms repre-
sent the approximated S term. It should be noted that the resistance term
does not always depend on one parameter only. A more complex resistance
term can be applied for certain cases, according to the selected failure
criteria. An example of this type of expression is a von Mises formulation
for an allowable stress envelope that connects failure stresses that may vary
for different directions. It should be emphasized that neither deterministic
nor probabilistic analysis can be performed if failure criteria are not formu-
lated for the case in hand. This formulation depends significantly on both
personal and organizational experience.
Once the approximate expression for the failure surface is formulated,
calculation of the probability of failure for each structural component can
be performed analytically (for simple cases) or numerically. The reliability
index can be found and the probability of failure determined using FORM
or SORM methods.
Figure 10.2 presents a computational flow chart of the Taylor series
expansion and probabilistic solution.

IV. Detailed Examples of the Taylor Series Expansion Method


Example 10.1
A cantilever beam of length L, rectangular cross section of width b, and
the height h is loaded by a force per unit length q. The selected failure
criterion is the beginning of the yield at the clamped end. The yield stress
of the material is S,. In Table 10.1, data of the PDFs, means, and standard
deviations of the random variables are shown.
The nominal stress of the beam (calculated with the mean values of the
parameters) is 3105 kg/cm2 and the traditional safety factor is 1.159.
For this structural component, the bending stress at the clamped end
(the load effects term in the failure function) is given analytically by

The resistance term is


Purchased from American Institute of Aeronautics and Astronautics

TAYLOR SERIES EXPANSION OF THE FAILURE SURFACE 141

I Define the structural )

-
system and components

I
i
I
1
Decide what are the
+ +
For each component
4
Form a List of variables
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

failure criteria influencing the suucture


I I

I
F l Decide on the random
1 variables of the load term I

random variables of densities, means and


I the resis

Determine nominator and


I denominator variables I
L- 4

Select evaluation point


I
and range for the
expansion
II
,Prepare a table of
deterministic cases

Check Xcr and Solve these cases


change EP and with an FE
range, if required

t
L
I Solve for hevector la)
1
I
-
-
Perform a second
iteration,
(if required)
7
I I Use avadable probabilistic I
cd-; 1
Results

Fig. 10.2 Flow chart of the computational procedure.


Purchased from American Institute of Aeronautics and Astronautics

142 G. MAYMON

Table 10.1 Data for Example 10.1

Standard
Physical Mean, deviation,
variable x, Units Distribution PCL, cl

4 XI kglcm Normal 1.15 0.0333


L x2 cm Normal 60 0.6
x3
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

b cm Normal 4 0.12
h x
4 cm Normal 1 0.03
s~ xs kg/cm2 Normal 3600 300

and the failure surface is

The problem was solved using four methods: 1) FORM solution of the
closed-form expression (10.13) using the Lagrange multiplier method; 2)
FORM, SORM, and Monte-Carlo solutions of the closed-form expression
(10.13) using the PROBAN program; 3) expansion of the failure surface
into a Taylor series and a FORM solution obtained by this approximate
expression using the Lagrange multiplier method; and 4) introduction of the
Taylor series approximate expression into the PROBAN and performing
FORM and SORM, as well as Monte-Carlo solutions.
Solution I : Transforming all of the basic random variables into standard
normal variables is performed using Eq. (8.27) and, therefore,

The failure surface in the standard normal space is


Purchased from American Institute of Aeronautics and Astronautics

TAYLOR SERIES EXPANSION OF THE FAILURE SURFACE 143

Creating the variable D according to Eq. (9.3) and using Eq. (9.4) yields
five equations. In addition to Eq. (10.15), these equations are solved for
the design point in the transformed space. Equation (10.13) is used to
obtain the design point in the basic variables space, and Eq. (9.2) is used
to calculate the reliability index.
The results are
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Solution 2: The following results were obtained using the PROBAN


program with the failure surface (10.15)

Using 5000 Monte Carlo simulations yields

PMonte
Carlo = 1.300; 1.261-1 .MI, 90% confidence interval

PfMonte Carlo = 0.0968; 0.08992-0.1037,90% confidence interval (10.17b)

Solution 3: The variables q and L are nominator type; b and h are


denominator type. The EP was selected as the mean plus or minus half-
standard deviation, as described in Table 10.2.
The number of required deterministic solutions is 9 (2 X 4 variables +
1). These are listed in Table 10.3. The solutions were performed with the
closed-form expression (10.11), which simulates for this example the finite

Table 10.2 Selection of EP, Example 10.1

EP A x EP - AXi EP + AX,
Purchased from American Institute of Aeronautics and Astronautics

G.MAYMON

Table 10.3 Nine deterministic solutions,


Example 10.1

4 L b h Sbend

1 1.16665 60.3 3.94 0.985 3329.1072


2 1.15 60.3 3.94 0.985 3281.5954
3 1.1833 60.3 3.94 0.985 3376.6190
4 1.16665 60.0 3.94 0.985 3296.0642
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

5 1.16665 60.6 3.94 0.985 3362.3151


6 1.16665 60.3 3.88 0.985 3380.5883
7 1.16665 60.3 4.00 0.985 3279.1706
8 1.16665 60.3 3.94 0.970 3432.8654
9 1.16665 60.3 3.94 1.000 3229.9830

element solution. The nine simulations are also presented in the last column
of Table 10.3. The [qmatrix corresponding to the cases in Table 10.3 is

Solution of Eq. (10.6) yields the following values for {a):

a0 = 3332.4332

a1 = 2853.561562; as = -21635.79

a2 = 110.4182; a6 = -36.03913

a3 = -84.51475; a7 = -709.367

a4= -6762.7467; as = -0.0004484


Purchased from American Institute of Aeronautics and Astronautics

TAYLOR SERIES EXPANSION OF THE FAILURE SURFACE 145

Table 10.4 A new selection of EP, Example 10.1

EP Ax, EP-AX, EP+AX,


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Introducing these values into Eq. (10.3) and solving the equations for the
critical values using Eq. (10.9) yields
q,, = 1.2326; LC,= 61.832
b,, = 3.88; h,, = -7.5 x lo6 (10.20)
The critical value of b is on the edge of the evaluation range and, therefore,
a new EP is selected for the width variable. This new EP and range are
summarized in Table 10.4.
The nine new deterministic cases to be solved, together with the calcu-
lated values of the vector Sbend, are shown in Table 10.5. The corresponding
[Y] is now

The solution for the values of {a}are


a. = 3380.6048
al = 2897.69069; a5 = -81.82272
Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON

Table 10.5 Nine new deterministic solutions,


Example 10.1

9 L b h Sbend
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

and the critical values calculated using Eqs. (10.3) and (10.19) are

q,, = 18.873; LC,= -14.9369


b,, = 5.448; h,, = 1.3307 (10.23)

which are all well outside the evaluation range.


Using Eq. (10.22) in Eq. (10.10), the Lagrange multiplier method, the
design point, the reliability index, and the probability of failure were calcu-
lated. The results are

Solution 4: The Taylor series for S [Eq. (10.3)] with the coefficients of
Eq. (10.22) were introduced into the PROBAN program. The results are
Purchased from American Institute of Aeronautics and Astronautics

TAYLOR SERIES EXPANSION OF THE FAILURE SURFACE 147

Table 10.6 Comparison of results for different solution methods, Example 10.1

Method Analysis P pf
-

1) Lagrange multiplier, Eq. (10.13) FORM


2) PROBAN, Eq. (10.13) FORM
SORM
Monte Carlo
3) Lagrange multiplier, Taylor FORM
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

expansion Eq. (10.10)


4) PROBAN, Taylor expansion FORM
Eq. (10.10) SORM
Monte Carlo

and using 10,000 Monte Carlo simulations

PMonte = 1.237; 1.210-1.265,90% confidence interval

PfMonte Carlo = 0.1081; 0.1030-0.1132,90% confidence interval


(10.25b)
The results of all calculation methods are tabulated in Table 10.6.

Example 10.2
A simple structural component that does not have a closed-form failure
function is treated in this example. The model, described in Fig. 10.3, is a
cantilever beam built of 26 E-glass/epoxy plies. The thickness of each ply
is h. They are all placed at an angle 8 to the main direction of the beam
and loaded by a uniform pressure p, which is random in its value but not
in its uniformity. It is assumed, for simplicity and without loss of generality,
that all plies are at one angle 9 and that the randomness is only between
different specimens. The length of the beam is L and its width is b.

Fig. 10.3 Cantilever beam model.


Purchased from American Institute of Aeronautics and Astronautics

148 G.MAYMON

The deterministic analysis of this model is solved with the ANSYS finite
element code, using the STIFF99 element, which is an eight-node layered
shell element. The data required to calculate the stresses of the deterministic
case (the input data to ANSYS) is as follows:
p = external applied pressure
h = thickness of one ply
n = number of plies
L = length of the beam
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

b = width of the beam


Ex = Young's modulus along the plies, x direction
Ey = Young's modulus perpendicular to the plies, y direction
E, = Young's modulus perpendicular to the plies, z direction
Gxy= shear modulus, directions x-y
G,, = shear modulus, directions x-z
Gy, = shear modulus, directions y-z
vxy,u,,, vyz= Poisson's ratios
It is good engineering practice to assume

It is assumed that L, n, vxy,and v,, are deterministic, with the following


values:
L = 18 cm; n = 26; vxy= 0.26; u,, = 0.49

Thus, the number of variables in the load effects term of the failure
surface is I = 6. The statistical characteristics of the remaining variables
are described in Table 10.7.
Many failure criteria for composite structures exist that are beyond the
scope of this discussion. By solving the nominal deterministic case and
comparing the obtained stresses and strains in tension, the compression
and shear along and perpendicular to the direction of the plies were deter-
mined. It was concluded that this particular structural component fails if
the tensile stress in the direction perpendicular to the fibers (tensile stress

Table 10.7 Data for Example 10.2

Xi Units Distribution Mean, pi Standard deviation, a,

XI= P kg/cm2 Norma1 0.14 0.00333


X2 = h cm Normal 0.019 0.001
X3 = Ex kg/cm2 Normal 0.386 X lo6 0.0338 X lo6
Xq = Ey kg/cm2 Normal 0.0827 X lo6 0.00124 X lo6
X5 = G, kg/cm2 Normal 0.0414 X lo6 0.00621 X lo6
x6= e deg Normal 45 1
Purchased from American Institute of Aeronautics and Astronautics

TAYLOR SERIES EXPANSION OF THE FAILURE SURFACE 149

Table 10.8 Selection of EP, Example 10.2

EP AX EP - AX; EP + AX,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

in the matrix material) is higher than the allowable tensile stress in the
matrix. It is, therefore, concluded that the resistance term of the failure
function is

where S;, is the allowable (failure) tensile stress in the y direction. From
available data, it is assumed that R is normally distributed with the following
mean and standard deviation:

The EP and range is selected according to the parameters defined in Ta-


ble 10.8.
The I deterministic cases to be solved, together with the tensile stress
perpendicular to the fibers S,,, are described in Table 10.9. In the last row
of Table 10.9 the deterministic case, which uses the average values of
all of the variables, is also described. The corresponding [Yl matrix is
Eq. (10.28).
Purchased from American Institute of Aeronautics and Astronautics

150 G.MAYMON

Table 10.9 Thirteen deterministic solutions, Example 10.2

P h Ex 4 Gv 0 s,
1 (EP)
2
3
4
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

5
6
7
8
9
10
11
12
13
Nominal

Solution for the coefficients {a}yields

a g = 321.84037

The critical values calculated using Eq. (10.3) and (10.29) are

p,, = 1.1825; h,, = 0.019677; Ex,, = 0.3533 X 106

Four of these values are within the evaluation range and, therefore, a
second calculation is required. This is performed by close observation of
the previous calculation. The second selected E P is shown in Table 10.10.
Purchased from American Institute of Aeronautics and Astronautics

TAYLOR SERIES EXPANSION OF THE FAILURE SURFACE 151

Table 10.10 A new selection of EP, Example 10.2

EP AX EP - AXi EP + AX,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

By solving an additional 13 deterministic cases with the ANSYS program


and constructing a new matrix [Yl, a second set of solutions is obtained
for {a}
a 0 = 320.57

al = 2247.112; a, = 900.29013

a6= 12.25; a12= -0.01392 (10.31)


In this case, the critical values of two of the variables Ex and G , are very
close to the end of the evaluation range. Nevertheless, the analysis is done
with these values. Introducing the coefficients Eq. (10.31) into Eq. (10.10),
using XF = S$ and applying the Lagrange multiplier method already de-
scribed, the following results for the design point, the reliability index, and
the probability of failure are obtained:
p* = 0.14087 kg/cm2
h* = 0.01775 cm
E,* = 0.380376 x lo6kg/cm2

s$]= 338.4427 kg/cm2


/3 = 1.7310766
Pf= (first order) = 0.0417191
Purchased from American Institute of Aeronautics and Astronautics

152 G.MAYMON

Table 10.11 Selection of four points evaluation method, Example 10.2

AX, EP-2AXi EP-AX, EP EP+4Xi EP-24Xi


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Table 10.12 Twenty-five deterministic solutions, Example 10.2


Purchased from American Institute of Aeronautics and Astronautics

TAYLOR SERIES EXPANSION OF THE FAILURE SURFACE 153

It can be seen that E,* is outside the evaluation range, whereas p*, E,Y,
G,y, and O* are close to the end of the range. Therefore, the {a)of Eq.
(10.31) are not a good set of coefficients for the approximated failure
function. It is, therefore, recommended to evaluate the failure function by
using four points around the EP. In Table 10.11, EP and ranges for four-
points evaluation are shown.
In this case 41 + 1 = 25 deterministic solutions are required. These are
shown in Table 10.12, together with the ANSYS solution for the tensile
stress perpendicular to the fibers.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The matrix [Ylthat corresponds to this evaluation has 25 rows and 13


columns and is Eq. 10.33.
Purchased from American Institute of Aeronautics and Astronautics

154 G.MAYMON

Solution of Eq. (10.6) yields the following results for the Taylor series coef-
ficients:
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

By introducing these values into the failure function (10.10) and applying
the Lagrange multiplier method, results for the design point, reliability
index, and probability of failure can be obtained. These are

O* = 45.38 deg

Pf(first order) = 0.033356 (10.35)

The same approximate closed-form expression was used in the PROBAN


program, to solve the problem by FORM, SORM, and Monte-Carlo simula-
tions. All of the results are summarized in Table 10.13. It can be seen from
the difference between the FORM and SORM results that the solved
example presents nonlinear behavior. Nevertheless, the approximate ex-
pression obtained by the Taylor series expansion enables the use of a
SORM analysis using an available program in which SORM solutions are
Purchased from American Institute of Aeronautics and Astronautics

TAYLOR SERIES EXPANSION OF THE FAILURE SURFACE 155

Table 10.13 Comparison of results for different solution methods, Example 10.2

Method Analysis P pf

Lagrange multiplier two-point FORM 1.7310766 0.0417191


evaluation
Lagrange multiplier four-point FORM 1.8336123 0.033356
evaluation
PROBAN, four-point evaluation FORM 1.834813 0.0332667
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

SORM 2.304510 0.010597


Monte Carlo 2.323 0.0101
(l0,OOO)

available. The SORM solution compares well to the Monte Carlo simu-
lation.

V. Summary
A Taylor series expansion method is used to create an approximate
closed-form expression for the failure function of a practical structure.
The expansion requires a solution of at least (21 + 1) deterministic cases,
where I is the number of random variables. These cases can be solved with
any analytical or numerical tool. The coefficients of the Taylor series are
obtained by a least-mean-square method.
The best evaluation point around which the function should be expanded
is the design point, which is not known in advance. Suggestions for the
practical selection of the EP are included, and an iterative process is sug-
gested.
Two examples demonstrate the application of the suggested process.
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

Chapter 11

Direct Calculation of the Probability of Failure


Using an Existing Finite Element Program
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Introduction

I N the previous chapter, a Taylor series expansion of the failure surface


around an arbitray selected E P was presented. The best E P for this type
of expansion is the design point, which is not known in advance. Therefore,
an iterative procedure was adopted.
Therefore, it is clear that advance knowledge of the design point is of
great benefit to the designer and the analyst. The modified joint probability
density function (MJPDF) method described in this chapter can provide
this knowledge. Moreover, it is shown that first-order probability analysis
of practical structures can be performed by using only a commercially
available finite element code, without the need for a probabilistic program.
In this chapter it is shown that the design point can be calculated by finding
a maximum to an objective function, which is obtained by modifying the
joint PDF of the relevant problem. The maximum is searched for in the
original physical random variable space X and not in the transformed space
U. In case of random fields rather than random variables, discretization of
these fields to random variables is performed according to the methods of
stochastic finite elements, e.g., Refs. 43 and 44. Thus, a solution can be
obtained using any available numerical algorithm that can find a maximum
to a given function. Preferred algorithms are those that do not require
calculation of numerical derivatives. The space in which this maximum is
sought is the allowable range of the physical random variables, which are
easily defined by the designer, and the finite element code is used as a tool
for numerical simulation. Once the design point is obtained, the reliability
index can be calculated, and the first-order probability of failure is obtained
directly. If second-order probability of failure is required, a Taylor series
expansion around the already-known design point is used.
All examples presented in this chapter were solved using ANSYS$5 which
is a large commercial finite element code without any modification in the
code. This code (like most commercial programs) contains a built-in optimi-
zation module$6 and this particular module is used to search for the maxi-
mum of the objective function without any modification of the finite ele-
ment code.
The advantage of using this approach is the possibility to use an existing
finite element code known to the design engineers, without modifying it
Purchased from American Institute of Aeronautics and Astronautics

158 G. MAYMON

and without using derivatives of the load effects term with respect to the
random variables. This approach can encourage engineers in industry to
use probabilistic structural analysis with existing tools until probabilistic
analysis modules are incorporated into commercially available finite ele-
ment programs.

11. MJPDF Method


Assume a typical basic failure function
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where XI is the basic load effects term, normally distributed with mean
pXland standard deviation uxl.Usually, XI is a function of the basic physical
random variables of the structure and the loading. In case of spatially
distributed random parameters, the load effects term also contains the
random variables that result from a discretization of random fields by any
of the methods described in the literature. X2is the basic random resistance
term, normally distributed with mean px2and standard deviation gX2.
If the load effects and the resistance terms are not normally distributed
(and may even be dependent), a transformation to independent normal
variables is available.34This is subject to the classical conditions that the
marginal CDF of the variable is continuous and that the transformed corre-
lation matrix of correlated variables is positively definite. The last condition
is satisfied in nearly all cases of practical interest.
Transformation into a standard normal space, e.g., Eq. (8.27), yields the
limit state function in the u space

Equation (11.2b) is described in Fig. 11.1.


By geometric considerations, the following quantities are calculated from
Fig. 11.1:

Px, - Px, "X,(PX, - Px, ) "X,(c~x, - P x , )


0 = qv; uf = ; U: = (11.3)
"xl + cx2 "x, + "2, dl+ "2,
Inverse transformation to the original X space yields the physical design
point
2
+ PX2"2,
x;= x; = P x , "x,
2
"XI + "2,
Purchased from American Institute of Aeronautics and Astronautics

USING AN EXISTING FINITE ELEMENT PROGRAM 159


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 11.1 Failure function in a standard normal space.

of XI, X2 is
The joint probability density function (JPDF) fxl,x2

The MJPDF P x , , ~is defined as the original JPDF, where the resistance
variable is replaced by the load effects variable

It is easily shown that this function has a maximum when

which is identical to Eq. (11.4). Thus, the design point is the point of
maximum of the MJPDF.
In realistic practical cases, the load effects term XI is calculated by the
required algorithm, i.e., a finite element code. This term contains many
Purchased from American Institute of Aeronautics and Astronautics

160 G.MAYMON

basic physical random variables (such as external loads, geometry, and


material properties). The JPDF contains n parts, which belong to these n
variables, plus one part, which originates from the resistance term. This
last part is the one for which the modification is performed: the variable
X2 in this part is replaced by the variable XI, which is the outcome of the
finite element computation for a specific required response, i.e., stress,
displacement, or moment.
A special case is obtained when the resistance term is a constant (a
threshold value problem). In this case, it can be easily shown that the design
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

point is obtained by finding the point that maximizes the JPDF of the load
effects term, while fulfilling the condition that the response is equal to the
threshold value. The response itself is an outcome of the finite element com-
putation.
It is important for the designer to be able to represent the JPDF for the
cases in hand. Three types of practical cases are possible:
1) All of the random variables are independent. In this case, the JPDF
of the random variables { X } is given by

where fXn(xn)is the probability density of the nth variable.


2) All of the random variables are normal, but there is a correlation
between some of them. In this case, the covariance matrix [R] is assumed
to be known, and the multivariate normal distribution function is used

where det[R] is the determinant of [R]. Note the following relation between
the covariance matrix and the correlation coefficients:

3) The random variables are dependent, and at least one of them is not
normally distributed. The Nataf47model for known marginal distributions
is used.
Purchased from American Institute of Aeronautics and Astronautics

USING AN EXISTING FINITE ELEMENT PROGRAM 161

Assume two dependent variables XI and X2with a correlation coefficient


pl,,. The basic variables X are transformed to standard normal variables
Z using
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where F( ) is the marginal cumulative distribution of the variable ( ) and


@ is the standard normal cumulative function. According to the Nataf
model, the JPDF is approximated by

where fx,(xi) is the marginal probability density of X i and & ( z l , z2,


is the biiariate normal PDF of the variables Z with correlation coefficient
~ . the case of two variables xl and x2
p ~ , , For

The coefficient p,,l,2 is obtained as a function of p l , ~using the relation

In Ref. 48, expressions for F are given for many combinations of distribu-
tions. The maximum error obtained in the JPDF using these tables is
usually less than 1%, although in very few combinations an error of 4.3%
was obtained.
Generalization of Eq. (11.1 1 ) yields

in which & ( Z , R o ) is a multivariate distribution factor similar to Eq. (11.9)


with covariance matrix Ro,whose elements are obtained by using the formu-
lations and tables published in Ref. 48.
Once the JPDF is formulated, the MJPDF is obtained according to the
procedure described earlier in this chapter, and the design point is computed
by the finite element code. The point of maximum of MJPDF is obtained
in the X space, whereas the reliability index is the distance between the
Purchased from American Institute of Aeronautics and Astronautics

162 G. MAYMON

design point in the transformed independent standard normal space u and


the origin of this space. Therefore, the reliability index cannot be calculated
straightforwardly, and some manipulations of the basic variables are re-
quired.
For the simple case of independent normally distributed variables, the
known design point {X*}is transformed to the standard normal space, using
Eq. (9.27) to obtain {u*}. The reliability index is obtained using
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

For independent variables that are not normally distributed, the design
point is transformed into the standard normal space using Eq. (11.10), and
then Eq. (11.15) is used to calculate the reliability index.
When the variables { X } are dependent and normally distributed with
a covariance matrix such as Eq. (11.9), they are first transformed into
independent normally distributed variables {Y},so that the covariance ma-
trix of Y is diagonal. This is performedz6 using

where [A]is an orthogonal matrix of which the columns are the eigenvectors
of the matrix [R].As Eq. (11.16) is a linear transformation, the mean of
any variable Y i is given by

where ai are the terms in the ith row of [AITand px,are the means of the
variables {X}.The standard deviations of { Y }are given by

Once the design point { X * ) is known, { Y * ) and its two first moments can
be calculated, and then {u*} is calculated using

Equation (11.15) is then used to calculate the reliability index.


When the variables are dependent and nonormally distributed, the results
obtained by the Nataf47 model are used. In Eq. (11.10) the variables {X}
with covariance matrix [R] were transformed into normal variables {Z}
with covariance matrix [Ro].To calculate the reliability index, these vari-
Purchased from American Institute of Aeronautics and Astronautics

USING AN EXISTING FINITE ELEMENT PROGRAM 163

ables should be transformed into independent standard normal variables


{Y). The following transformation between {u) and (2)is performed:
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

It can be shown that [To]= [Lo]-',where [Lo]is the lower triangular part
of the Choleski decomposition of [Ro].For a known [Ro],the [Lo]can be
calculated using [R,] = [Lo][Lo]-'.
Once [ Y ]is known, Eqs. (11.15) and (11.17-11.19) can be used to com-
plete the analysis.
The preceding discussion of the MJPDF method may seem complicated,
but inasmuch as this method is not covered in the literature (except for Ref.
49) it is felt that some theoretical background is required. The numerical
examples in the following section demonstrate the use of the method.

111. Numerical Examples


All of the numerical examples presented in this chapter were solved
using the ANSYS finite element program without any modification. ANSYS
is used for calculation of the load effects term (denoted in this section by
SFEto emphasize it was computed by a finite element code). Expressions
for the JPDF and MJPDF were written and introduced into the proper
modules of ANSYS.
The optimization module of ANSYS was used to find the maximum
point of the relevant function. This module creates an objective function
approximation in the form of a quadratic equation, including cross-product
terms. The constrained problem is then converted into an unconstrained
one, using penalty functions. The loop includes a search for a minimum of
the objective function, the determination of the next trial, and the calcula-
tion of the next step. Optimization looping continues until a predetermined
convergence criterion is met. The algorithm was written originally for design
optimization and, therefore, a minimum is sought. To use the same subrou-
tines in the search for maximum value of the MJPDF, the objective function
introduced into the program is the inverse of the MJPDF.
Although ANSYS has several mathematical functions in its library, the
error function is not included. Therefore, it was necessary, in some cases,
to describe the error function as an approximate polynom. This is a disad-
vantage that does not originate from the MJPDF method and can be over-
come in the future by introducing the error function into the program
library. In Appendix C, input files of the ANSYS program for some of the
examples are presented. The reader should note that this input is not
suitable for other finite element programs, and it is presented only to show
the simplicity of the application of the MJPDF method. ANSYS 4.2 version
was used for the computations. In more updated versions of ANSYS (in
1996, version 5.3 was available) the procedures for optimization are slightly
Purchased from American Institute of Aeronautics and Astronautics

164 G.MAYMON

different but this does not change the major processes used by the
MJPDF method.
In the following subsections, details of the numerical examples are pre-
sented.

A. All Variables Uncorrelated and Normally Distributed


Example 11.1
A bar of cross section A is loaded by an axial force F. The bar has an
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

yield stress S,. Failure occurs when the stress in the bar is equal to or
higher than the yield stress. All of the variables are independent. Data
on the three variables are given in Table 11.1. The closed-form failure
function is
g = S y - (FIA) 5 0 (11.21)
The mechanical transformation is known in this case and is FIA. Assuming
that this transformation is unknown, the load effects terms is SFE,the axial
stress calculated by the finite element program, and is a function of the
first variable F and the second variable A. The failure function is, thus,

The JPDF of the three variables is

The MJPDF is obtained by replacing Sy with SFE

Table 11.1 Data for Example 11.1

Variable Distribution Mean Standard deviation

F Normal p, = 1000 a, = 33
A Normal PA= 2 u~ = 0.1
SY Normal psy= 600 us?= 30
Purchased from American Institute of Aeronautics and Astronautics

USING AN EXISTING FINITE ELEMENT PROGRAM 165

Table 11.2 Results for Example 11.1

ANSYS PROBAN Error, %


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

and the maximum of Eq. (11.24) is sought by running the proper ANSYS
case. Input files for ANSYS are described in Appendix C, Sec. I. The
variable space in which this maximum is searched for is selected as

mean = 1000 5 F r 1099 = mean + 3 standard deviations


mean - 3 standard deviations = 1.7 5 A 5 2 = mean (11.25)

As this example has a closed-form failure expression, it was also possible


to solve it using the PROBAN program. In Table 11.2, results of the design
point, reliability index, and probability of failure are presented for both
the MJPDF method and the direct solution by PROBAN. The design point
in the standard normal space u* was calculated using Eq. (11.27). The
reliability index was calculated using Eq. (11.25). The first-order probability
of failure presented in Table 11.2 is obtained using Eq. (8.13).

Example 11.2
A cantilever beam of length L is subjected to a tip force F. The cross
section of the beam is rectangular, with height H and area moment of
inertia I. The yield stress of the beam material is S,. H = 2 is deterministic,
whereas all of the other variables are independent and normally distributed.
Data are summarized in Table 11.3. Failure is defined when the bending
stress in the clamped end is equal to or higher than the yield stress. The
closed-form failure function can be expressed using simple beam theory as

Table 11.3 Data for Example 11.2

Variable Distribution Mean Standard deviation


Normal p~ = 1000 a~= 33
Normal PI =2 al = 0.1
Normal p~ = 50 a, = 2
Normal psv= 33,000 as"= 1000
Purchased from American Institute of Aeronautics and Astronautics

166 G. MAYMON

If no closed-form expression is known, the failure function is expressed as

where SFE is the bending moment at the clamped end, as calculated by a


finite element program.
The JPDF is given by the following expression:
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The MJPDF is obtained by replacing Sy with SFE

A finite element beam model was run in the ANSYS program, using the
optimization module where F, L, and Zwere changed and SFEwas calculated,
until a maximum was found for Eq. (11.29). The input file for ANSYS is
described in Appendix C, Sec. 11. Results are summarized in Table 11.4,
together with PROBAN results obtained by using Eq. (11.26). The design
point in the standard normal space u* was calculated in the same way as
for example 1.

Table 11.4 Results for Example 11.2

ANSYS PROBAN Error, %


Purchased from American Institute of Aeronautics and Astronautics

USING AN EXISTING FINITE ELEMENT PROGRAM 167

B. Normally Distributed Correlated Variables


Example 11.3
Example 1 is repeated with the same data given in Table 11.1, but with
a correlation coefficient p = 0.5 between the force F and the area A. The
JPDF is given by
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The MJPDF is

where SFEreplaced S,. The design point was obtained using the ANSYS
program to find F and A that maximize Eq. (11.31). The input file for
ANSYS is described in Appendix C, Sec. 111. Results are summarized in
Table 11.5, and compared to PROBAN results. The u* point was calculated
as follows.

Table 11.5 Results for Example 11.3

ANSYS PROBAN Error, %


Purchased from American Institute of Aeronautics and Astronautics

168 G.MAYMON

The covariance matrix [ R ] is

The use of a 2 X 2 matrix is sufficient, as there is no correlation with


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

S,. Therefore,

The eigenvalues of [ R ] are

The matrix [ A ] ,which is the eigenvector matrix of [ R ] ,and its transpose


[ A ] [see Eq. (11.16)] are

Therefore,

Introducing F* and A* from Table 11.5 into Eq. (11.35) yields

Applying Eqs. (11.16) and (11.17) yields


Purchased from American Institute of Aeronautics and Astronautics

USING AN EXISTING FINITE ELEMENT PROGRAM 169

Finally, applying Eq. (11.19) yields the values of u*, which are given in
Table 11.5.

C. Nonnormal Dependent Variables


Example 11.4
A bar of cross-section A and yield stress S,, both normally distributed,
is subjected to an axial force T , which has a three-parameter Weibull
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

distribution, with a minimum value of 900. There is a correlation between


T and A, with a correlation coefficient 0.4. The data are summarized in
Table 11.6.
The three-parameter Weibull distribution has the following CDF:

F*(X) = 1 - exp [ (-)


-
X - E
k -E
'1
and the following PDF:

where x 2 E , p > 1, and k > E 2 0, and E , /3, and k are given constants
of the specific problem. The mean and the standard deviation of the Weibull
distribution is given by

Table 11.6 Data for Example 11.4

Variablea Distribution Mean Standard deviation

T Weibullb p~ = 1000 LTT = 50

A Normal PA = 2 u~ = 0.1
s~ Normal psY= 700 usY = 33

"Coefficient p,, = 0.4. bMinimum value of T is 900.


Purchased from American Institute of Aeronautics and Astronautics

170 G.MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Tensile Force (kg)


Fig. 11.2 PDF of the force T.

Sometimes F(x) and f (x) are described as follows:

Introducing the variable T (the axial force) into (11.37d) and using the
data of Table 11.6 and Eq. (11.37c), the following PDF is obtained for T:

where p = 2.1013491, S = 4.8587 X and y = 900. These values


correspond to a Weibull distribution with mean p~ = 1000, standard devia-
tion UT = 50, and minimum value y = 900, as given in Table 11.6. This
function is described in Fig. 11.2.
The CDF for the Weibull distribution is

To write the JPDF for this case according to Eq. (11.11), Z1 and Z2,which
correspond to T and A, respectively, are to be formulated. Using Eq. (11.16)
the following is obtained:
Purchased from American Institute of Aeronautics and Astronautics

USING AN EXISTING FINITE ELEMENT PROGRAM 171


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

1000 1050 1100 1150 1200


Tensile Force (kg)
Fig. 11.3 Polynomial solution of Z1 compared to the exact solution.

To express Z1 as a function of T inside a numerical program, a function


library with the error function is required. Because the ANSYS library
does not include this function, a polynomial approximation is formed, and
the first of the two equations (11.39) is replaced by

In Fig. 11.3, the polynom solution of Z , is compared with the exact solution.
The matrix of the variance coefficients between T and A is

Using the tables of Ref. 48, the following is obtained:

and, therefore,
Purchased from American Institute of Aeronautics and Astronautics

172 G.MAYMON

The JDPF of the two dependent variables T and A, according to Eq.


(11.11), is
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where $( ) is the standard normal probability density function of ( ) and


Ziis given as a function of Xiby Eqs. (11.39).
The JPDF of the three-variables system is given by

and the MJPDF is given by

The maximum of this function is found using the ANSYS program. The
input file for ANSYS is described in Appendix C, Sec. IV. The values of
T* and A * are described in Table 11.7. Using these values in Eqs. (11.39),
the values of Z; and Z; are

Table 11.7 Results for Example 11.4

ANSYS PROBAN Error, %


Purchased from American Institute of Aeronautics and Astronautics

USING AN EXISTING FINITE ELEMENT PROGRAM 173

To use Eq. (11.20) to find {u), it is necessary to calculate [To].This is


performed using the matrix [Ro]= [Lo][Lo]-', [To]= [Lo]-'.It can be
found that
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Introducing the second equation of Eq. (11.46) and the values of {Z*)from
Eq. (11.45), and taking into account that S, has an independent normal
distribution, the following is obtained for the vector {u*):

1.8509317
{u*) = -2.6866055
-2.9513636
I
These values enable calculation of the reliability index and the first-order
probability of failure can be computed. All of the results are summarized
in Table 11.7.
A high reliability index indicates that the design point is at the far tail
of the distribution and, therefore, the error in the probability of failure is
high relative to the error in the reliability index. There is no physical
significance to this error, as can be concluded when inspection of the numeri-
cal values is performed.

D. Threshold Value Problem


Example 11.5
A bar of cross section A is loaded by an axial force F. The data for A
and F is identical to the values presented in Table 11.1, example 1. The
bar is assumed to fail when the stress in it reaches the value of 570 (deter-
ministic).
The failure function for this case is
g = 570 - (FIA) 5 0 (11.48a)
If the stress in the bar FIA is not known as a closed-form expression, and
the failure function is

where is a function of A and F. This is a threshold value problem, and


the design point is found by determining the values of A and F that maximize
Purchased from American Institute of Aeronautics and Astronautics

174 G. MAYMON

Table 11.8 Results for Example 11.5

ANSYS PROBAN Error, %


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

the JPDF, subject to the condition that the stress is 570. The JPDF is given by

The input file is described in Appendix C, Sec. V, and is similar to that


described in Appendix C, Sec. I, with the proper modifications. As the
ANSYS optimization module requires a range of treated variables, the
problem was solved defining a small range between 569.6 and 570 for the
threshold value. The results are summarized in Table 11.8.

E. Random Field, Correlated, Normally Distributed Variables


Example 11.6
A beam of length L = 32 is clamped on both sides and has a random
flexural rigidity EI with mean p~ = 1.125 X lo6 and standard deviation
UE = 225,000. On the beam, a distributed load per unit length P i s applied,
with mean pp = 8 and standard deviation u p = 2.4. There is a correlation
in the force (per unit length) P between any two points AX apart, with a
correlation coefficient given by

where a, is a nondimensional correlation length and a, = 0.25 is assumed


for this example. Therefore, the force P is a random field rather than a
random variable.
The beam is assumed to fail when the bending moment at the clamped
end is equal to or higher than 1100. The beam is shown in the upper part
of Fig. 11.4. This example is a special case of the example solved in Ref.
44 using the stochastic finite element method. In Ref. 44, the flexural ridigity
EI is also assumed to have a correlation between any two points AX apart
and is, therefore, a random field. In this example, EZ is assumed to be a
random variable.
Using the same technique as in the application of the stochastic finite
element, the beam was divided into 32 finite elements and into 4 stochastic
Purchased from American Institute of Aeronautics and Astronautics

USING AN EXISTING FINITE ELEMENT PROGRAM 175


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Fig. 11.4 Original and discretized beam.

elements. Therefore, the case solved is the one described in the lower part
of Fig. 11.4.
There are many ways to discretize the random field into several random
variables. In this example, the value of the force at the center of the
stochastic element was selected to represent its value along the whole
element. This method is less accurate when the element is too large or
when changes in the value of the force over the selected element are large.
Sometimes, the spatial average of the field along the element is used, and
the load is applied at the center of gravity of the loading. It has been shown
in the literature that the length of the stochastic element should be between
a quarter and a half of the correlation length. According to this rule, at
least eight stochastic elements should be selected. For simplicity of the
demonstrated example, the length of the stochastic finite element is selected
to be equal to the correlation length; therefore, a difference between the
results of this example and Ref. 44 is expected. Nevertheless, as the case
presented in Fig. 11.4 can be solved explicitly, solutions using the MJPDF
method and the PROBAN program can be compared.
The covariance matrix for discrete values Piis given by
Purchased from American Institute of Aeronautics and Astronautics

176 G. MAYMON

Introducing the data just given, the covariance matrix is

1 0.367879441 0.135335283 0.049787068


1 0.367879441 0.135335283
[R] = 5.76 .
1 0.367879441
symmetric 1
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The determinant of the covariance matrix is 3.72362293476, and the inverse

1.156517643 -0.42549064 4.0997267 X lo-" 2.9021289 X lo-''


1.31303285 -0.42549064 4.0997267 X lo-''
1.31303285 -0.42549064
symmetric 1.156517643

The expression within the exponent of Eq. (11.9) is, therefore,


Purchased from American Institute of Aeronautics and Astronautics

USING AN EXISTING FINITE ELEMENT PROGRAM 177

Table 11.9 Results for Example 11.6

ANSYS PROBAN Error, %

P
P2
; 12.254
13.879
u; = 0.08543655
u,* = -0.36681643
12.29
13.879
0.29
0.20
p: 12.4374 uf = -0.75395417 12.4374 0.20
p: 10.0062 u: = 2.68164 10.0353 0.29
EZ* 1,123,500 ug = -0.0066667 1,125,000 0.13
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

P 2.81097 2.8250 0.5


pi 0.00246962 0.00236403 4.5

Because Eq. (11.51) is proportional to the JPDF of the four discrete vari-
ables Pi,an expression that is proportional to the total JPDF of the system
is obtained by multiplying the JPDF by the PDF of the rigidity EZ. An
expression proportional to the JPDF of the whole system is, therefore,

The ANSYS solution, based on 32 beam elements, was solved for the
bending moment at the clamp. The input file is described in Appendix C,
Sec. VI. The optimization module was run to find the values of the parame-
ters Piand EI that maximize Eq. (11.52), subject to the condition that the
moment is the given threshold value of 1100. Results are described in
Table 11.9.
To compute the reliability index, the matrix [A] of the eigenvectors of
the matrix [R] is first calculated. The results are
Purchased from American Institute of Aeronautics and Astronautics

178 G.MAYMON

Using Eqs. (11.16-1 l.l8), the following vectors can be obtained:


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

from which the vector {ui*}can be obtained using Eq. (11.19). Vector
components are summarized in Table 11.9. As EZ is uncorrelated, the
normally distributed variable, uE can be calculated directly from the physical
values appearing in the same table.

IV. Summary
A method for a direct calculation of the first-order probability of failure
of a practical structure using an existing finite element program is described.
The computation is done by evaluating an MJPDF and finding the maxi-
mum of this function. It is shown that this maximum point is the design
point, the most probable point of failure.
The computation is based on the existing design optimization module,
which is included in the commercial finite element code for quite a differ-
ent application.
Once the design point is found, the calculation of the first-order probabil-
ity of failure is straightforward. If second-order probability of failure is
required, the known design point can be used as an evaluation point for a
Taylor series expansion, for use as an approximate closed-form expression
in one of the probabilistic structures computer programs.
A variety of examples demonstrate the application of the method for
different kinds of practical structures, with different distribution functions
of the random variables and different relations between these variables.
These examples are solved using the ANSYS finite element program, and
results are successfully compared with those calculated using the PROBAN
program. Input files for the ANSYS program are presented in Appendix C.
The ability to use an existing finite element program may encourage the
application of probabilistic methods in the industry and encourage the
suppliers of finite element programs to include a probabilistic module in
the commercially available codes.
Purchased from American Institute of Aeronautics and Astronautics

Chapter 12

Probability of Failure of Dynamic Systems


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Introduction

I N Chapter 5, the response of a deterministic structure to stationary


random excitation was described and demonstrated for SDOF, MDOF,
and continuous systems. It was assumed that parameters of the structure
are known constants, and the PSD was calculated as a function of the
structure parameters and the PSD of the external excitation. The calculated
response can be a displacement at a certain point or, using the stress modes
approach, a certain stress in the structure.
In this chapter, the solution is extended to a nondeterministic structure,
for which certain structural parameters, i.e., dimensions and material prop-
erties, are also treated as random variables. If a failure criterion is defined
(such as a maximum allowable deflection or stress, which also may be
considered as a random variable) the probability of failure can be calculated
using the methods described in Chapters 8-11.
A stationary Gaussian excitation process with zero mean is a common
case for many practical structures. This case, which results in a stationary
Gaussian response process, can be solved by explicit expressions, so that
the phenomena can be demonstrated and, therefore, clearly explained. In
the following paragraphs, stationary Gaussian processes with zero means
are assumed.

11. Statistical Behavior of a Stationary Gaussian Process

Assume that a certain structure response S(t) (which may be a displace-


ment, a velocity, or a stress) can be described by a stationary Gaussian
process with zero mean and a PSD function G s ( o )in the positive frequency
domain 0 5 o 5 m. The absolute value of the maximum of this process
during a period 7 is defined by the variable S,

S, = max IS(t)l (12.1)


7

S ( t ) can be a function of { z ) ,a vector of structural random variables. Assume


that s is a threshold value (which may also be random and, therefore, is
a part of the vector ( 2 ) ) so that when the response is higher than s, the
structure fails. In Refs. 25 and 50 the conditional probability of failure (for
Purchased from American Institute of Aeronautics and Astronautics

180 G.MAYMON

given vector {z}) is found to be

where Pfo(z)is the conditioned probability that, at time t = 0, S(t) is higher


than s, given the values of {z}, and v(z) is the rate (times per unit time)
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

of the response process crossing the value s upward. It was also shown5' that

Pfo(z) = 1 - exp
( a
--

and A,' Al [which does not appear in Eq. (12.3)], and A2 are called spectral
moments of the process and are defined as

A,,,= 1; omGs(o)do; m=0,1,2 (12.4)

Thus, Eqs. (12.2-12.4) describe the conditioned probability that the process
S(t) has a value higher than a prescribed values, which is defined as a failure.
It can be seen that A. is the mean square value of the process, i.e., Eq.
(5.13), and A2 is the mean square value of the time derivative of the process,
i.e., Eq. (5.47). Spectral moment Al does not have a direct physical meaning.
The spectral moments are functions of G,(o), the PSD of the response,
which is a function of the PSD of the excitation and other structural parame-
ters. Therefore, the spectral moments are functions of all of the components
of the random vector {z} of the structure, including the excitation.
Substituting Eq. (12.3) into Eq. (12.2) yields

where Fs, is the CDF of the maximum values ST.


It was also shown that the zero crossing rate (the rate at which the process
crosses the time axis) is
Purchased from American Institute of Aeronautics and Astronautics

PROBABILITY OF FAILURE OF DYNAMIC SYSTEMS 181

The number of zero crossings at a period T is

and the number of upward zero crossings (crossing the zero axis in the
upward direction) is

The rate of upward crossings of the level s is given by Eq. (12.3).


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

In the evaluation of Eq. (12.2), some limiting assumptions were


One of the assumptions is a Poisson distribution for the maximum values.
Using Ref. 51 some assumptions are released and an empirical coefficient
q,, obtained by comparing results to numerical simulations, was introduced.
In Ref. 52, following Ref. 51, a less conservative expression for Fs, was ob-
tained,

Pf(z)= 1 -[I - exp (&)I


x exp [ -uOT - 1 - exp(-* . q, . s i f i ) ]
exp(-s2/2Ao) - 1
where

The PDF of ST,fs,(s), is obtained by

where

a2 - a2 exp(a9)
fs, = exp
exp(ao4) - 1
Purchased from American Institute of Aeronautics and Astronautics

182 G.MAYMON

and

In Fig. 12.1 the conditioned probability of failure described by Eqs. (12.5)


and (12.10) are shown for different periods for the following set of spec-
tral moments:
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Figure 12.1 should be read as follows: After, say, 1000 s, the probability
that there is a maximum value higher than 0.4 is 1 using either Eq. (12.5)
or Eq. (12.10). The probability of having a maximum value higher than 0.5
is 0.14 [using Eq. (12.5)] and 0.1 [using Eq. (12.10)]. It is evident that for
short durations the probability of having a maximum value that is higher
than a relatively low threshold is small, and this probability increases as
the process continues in time.
In Fig. 12.2, the CDF functions Fs, [Eq. (12.9)], the maximum values of
ST, are shown for different periods. As time increases the probability that
the maximum value of the process is lower than a certain threshold in-
creases.
In Figure 12.3 the PDF [Eq. (12.11)] is shown. As time increases more
and more maximum values of the process are included in a higher range
of threshold s.

0 0.1 0.2 0.3 0.4 0.5 0.6


Threshold s
Fig. 12.1 Conditional probability of failure.
Purchased from American Institute of Aeronautics and Astronautics

PROBABILITY OF FAILURE OF DYNAMIC SYSTEMS 183


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7


Threshold s
Fig. 12.2 CDF of the maximum values of S,.

0 0.1 0.2 0.3 0.4 0.5 0.6


Threshold s
Fig. 12.3 PDF of the maximum values of ST.
Purchased from American Institute of Aeronautics and Astronautics

184 G. MAYMON

111. Spectral Moments of a Deterministic MDOF System


To demonstrate the computation process of the spectral moments, an
MDOF system subjected to random excitation is considered. As shown in
Chapter 5, this type of system may represent a discretized continuous
system. The PSD of the response is obtained using Eq. (5.20), which is
repeated here
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The spectral moments of this system can be calculated using

The first expression is the mean square of the displacement of the jth DOF
and the last expression is the mean square of the velocity of this DOF.
As an example the two-DOF system described in Fig. 5.2 is treated. In
Eqs. (5.30) and (5.31), Sxl and S,, are expressed, respectively. Therefore,

- 0.24390.2 1.0
oRe[%(o)?ii.(o)] do} (12.14b)
Purchased from American Institute of Aeronautics and Astronautics

PROBABILITY OF FAILURE OF DYNAMIC SYSTEMS 185

Table 12.1 Results for some dynamic properties

Mass 1 ( j = 1) Mass 2 ( j = 2 )
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Using the expressions of Appendix A for the evaluation of the integrals


in the last six equations, the results described in Table 12.1 are calculated.
Once the values in Table 12.1 are known, Eqs. (12.5) and/or (12.10) can
be solved.

IV. Probability of Threshold Crossing (Failure)


In Eqs. (12.5) and (12.10), a conditioned probability of threshold crossing
is formulated. This probability is conditioned on the known values of the
structural random variable vector {z). In Ref. 53 it was suggested that
the nonconditioned probability can be calculated using the following limit
(failure) function:

@-' is the inverse standard normal distribution function and u n +is~ a stan-
dard normal variable added to the structural random variables. In Ref. 53
it was shown that when the structural random variables { z ) are transferred
into { u ) variables [using, for example, Eq. (8.27)], the design point is ob-
tained by finding the minimum distance between the origin of the u space
Purchased from American Institute of Aeronautics and Astronautics

186 G. MAYMON

and the multidimensional surface

Thus, the tools described in Chapters 8-11 can be adopted using Eq. (12.15)
as a failure function.
The described method of solution is possible only if the following two
conditions exist:
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

1 ) The boundary conditions of the structure are deterministic and do


not change between various specimens of the same structural design.
2 ) The parameters of the structure are random variables and not random
fields. This is true when parameters are changed between specimens and
not within specimens. For instance, if a thickness is a random variable,
this indicates that two specimens may have different thicknesses within a
prescribed statistical distribution, but the thickness does not change within
one given specimen.
These two conditions mean that the mode shapes are deterministic and
identical for all specimens, although the natural frequency of each specimen
may be random. Randomness in mode shapes is not treated herein.
When only a small number of structural random variables exist, the
following simple method can be used to find the probability that a displace-
ment (or a stress) is higher than a given random threshold:
1 ) Express the PSD of the response as a function of the structural parame-
ters and the PSD of the external excitation.
2 ) Express the spectral moments of the response.
3) Select a set of ranges for the random variables vector {z}.This selection
should be in a reasonable range of each z , according to its prescribed distri-
bution.
4) Calculate P f ( z ) using Eq. (12.5) or Eq. (12.10).
5 ) Compute k = W 1 [ P f ( z ) ] .
6) Using Eq. (12.16), set u,+~= k.
7 ) For each selected z, find the corresponding u using Eq. (8.27).
8 ) Calculate P = g u : + u? + + u: + ~ 2 , ~ .
9) Select another set of { z } values.
10) Repeat steps 5-8.
11) Find the minimum 6 for all of the possible sets of {z).
Such a calculation may be very long if there are many random variables
{ z } and it is, therefore, practical only for simple cases where this number
is small. The computation can be automated using a simple computer
program. The method is really a brute force scanning of the whole space
of { z } values. A flow chart for this type of calculation is shown in Fig. 12.4.
When a probabilistic computer program is available, this procedure can
be avoided.
It should be noted that steps 1 and 2 include the natural frequencies of
the system. When a closed-form expression for the frequencies exists, the
rest of the calculation is simple. When this is not the case, as often occurs
in practical structures, an approximate closed-form expression can be
Purchased from American Institute of Aeronautics and Astronautics

-
PROBABILITY OF FAILURE OF DYNAMIC SYSTEMS 187

Create Approximate 1. Express PSD of the


closed-form expression response as a function of all
for the natural the random variables

2. Express spectral
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

+ I

I
moments as fhnction of all
the random variables

3. Select a set of random


I

I
4. Calculate Pf(z) using

I (12.5) or (12.10) I
i
5. Calculate ~=CJ-'[P~
(z)]

7. From selected {z} find

8. Calculate P and store

previous one, replace stored

C e r I Exit

Fig. 12.4 Flow chart for scanning for minimum P.


Purchased from American Institute of Aeronautics and Astronautics

188 G. MAYMON

formulated using a Taylor series expansion. This will be demonstrated later


in the section.

Example 12.1
A simply supported beam of length L, circular cross section of diameter
d, made of a material of mass density p, with Young modulus E, and modal
dampings 4 is subjected to a stationary random load per unit length with
a PSD function of a one-sided white noise of magnitude So. Only the first
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

three modes are considered. The nominal data are

p = ylg = 7.959 X kg s2/cm4

4 = 0.01 for all j (12.17)

The mode shapes of the beam are1'

and the natural frequencies are18

where the cross-sectional moment of inertia I was replaced by nd4/64 and


the cross-sectional area A was replaced by nd2/4. Equation (12.19) yields
the following results for the first three natural frequencies:

ol= 258.6566 radls (41.17 Hz)

o2= 1034.6264 rad/s (164.67 Hz)

o3= 2327.9099 rad/s (370.50 Hz)


Purchased from American Institute of Aeronautics and Astronautics

PROBABILITY OF FAILURE OF DYNAMIC SYSTEMS 189

Using Eqs. (5.36) and (5.37), the following result is obtained for SQ,Qk:

It can be seen from the zero terms in Eq. (12.21) that the second mode,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

which is antisymmetric, does not participate in the response, because the


random loading is symmetric.
Substituting Eq. (12.21) into Eq. (5.39) yields the PSD function of the
response at point 6

For the center of the beam, 6 = 0.5, +l(e = 0.5) = 1, and &([ = 0.5) =
- 1; thus,

using Eq. (12.4) and the integrals in Appendix A, the first spectral moment
can be obtained,

Substituting the nominal values (12.17) into Eq. (12.24) shows that the
second term of Eq. (12.24), which is the contribution of the third mode, is
7.61 X times the first term, and the third term, which is the interaction
between the first and the third resonances, is 1.58 X times the first
Purchased from American Institute of Aeronautics and Astronautics

190 G. MAYMON

term. Thus, the response behavior at the center of the beam is governed
by the first mode only. Therefore,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The zero-crossing rate (12.6) is

and

where

Substituting the expression for the frequency (12.19) into Eqs. (12.25) and
these equations into Eq. (12.5) yields

where zl is the failure threshold, i.e., when the response of the center of
the beam is higher than zlthe beam fails.
In practical cases the frequency is usually not known as an explicit expres-
sion. Evaluation of an approximate closed-form expression for the fre-
quency using a Taylor series expansion is now demonstrated.
Purchased from American Institute of Aeronautics and Astronautics

PROBABILITY OF FAILURE OF DYNAMIC SYSTEMS 191

Table 12.2 Seventeen deterministic solutions

Case E

1 2.10 x lo6
2 1.96 X lo6
3 2.03 X lo6
4 2.17 X lo6
5 2.24 X lo6
6 2.10 X lo6
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

7 2.10 x lo6
8 2.10 x lo6
9 2.10 x lo6
10 2.10 x lo6
11 2.10 x lo6
12 2.10 x lo6
13 2.10 X lo6
14 2.10 X lo6
15 2.10 X lo6
16 2.10 X lo6
17 2.10 X lo6

The resonance frequency, which is a function of E, d, p, and L, is evaluated


into the following Taylor series around the nominal point p ~p d, , pp,and pL:

The four-points evaluation method described in Chapter 11 is used; there-


fore, 4 X 4 + 1 = 17 deterministic solutions are required. These 17 cases
are described in Table 12.2. The column w,,,, in the table describes the
deterministic frequencies which can be calculated for a practical structure
by running a finite element program.
The matrix of differences [Y] [see Eq. (10.5)J for the cases presented in
Table 12.2 is shown in Table 12.3.
Using

in which { w } is the column w,,,, of Table 12.2 yields the best coefficients { a ) ,
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Table 12.3 Matrix of Differences [ Y ]


Purchased from American Institute of Aeronautics and Astronautics

PROBABILITY OF FAILURE OF DYNAMIC SYSTEMS 193

Table 12.4 Data for the random variables

Standard Range, from


Variable Distribution Mean deviation -30 to +3u

ZI Normal 0.4 0.0133 0.36-0.44


E Normal 2.1 x lo6 70000 1.89 X lo6-2.31 X lo6
l Normal 0.01 0.000333 0.009-0.011
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Substituting these coefficients into Eq. (12.29) yields the approximate


closed-form expression for the frequency. The values obtained using this
expression are shown in the last column of Table 12.2. Very good compari-
son can be seen between the calculated and approximated values of the fre-
quency.
To demonstrate the calculation of the probability of failure, it was as-
sumed, without any loss of generality, that only three parameters of the
problem are random, the threshold value z l , Young modulus E, and the
modal damping coefficient & . The distributions of these variables and their
ranges are shown in Table 12.4.
The other structural parameters of the problem, So, d, L, and p, were
considered deterministic and the nominal values given in Eq. (12.17) were
assumed. The calculations are done for a period of t = 1 s, using the
approximate closed-form expression (12.29) for the resonance frequency,
with the coefficients (12.31). The scanning method described in Fig. 12.4
as well as the CALREL computer program were used. Results are shown
in Table 12.5.
The results show that for the given random variables, the most probable
point of failure occurs when the threshold value is 0.3945 (less than the
mean), Young modulus is 2.078 X lo6 (less than the mean), and the modal
damping coefficient is 0.00993 (less than the mean). With this combination,
the first-order probability of failure (defined as an amplitude higher than
the threshold) is 1.71%, and the second-order probability of failure is 1.75%.

Table 12.5 Results for the numerical example


Scanning CALREL
method program
Threshold value
Young modulus
Damping coefficient
Reliability index (FORM)
Probability of failure (FORM)
Reliability index (SORM)
Probability of failure (SORM)
Purchased from American Institute of Aeronautics and Astronautics

194 G. MAYMON

V. Summary
A calculation process for the probability of failure of a vibrating structure
is described. Failure is defined as the case where a certain response, i.e.,
displacement or stress, is higher than a certain given threshold.
Expressions for this probability of failure are shown for a stationary
Gaussian response process with zero mean, which is a common and useful
case in structural dynamics. The analysis is limited to cases for which the
mode shapes are deterministic. Nevertheless, the natural frequencies may
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

be random.
The spectral moments are defined, and their calculation is demonstrated.
Inasmuch as these moments are functions of the natural frequencies, a
Taylor series expansion method for creating an explicit expression for these
frequencies is also described.
By adding an extra standard normal variable, a failure function is built.
The probability of having a response higher than a prescribed threshold
can then be calculated using techniques described in preceding chapters,
or using one of the probabilistic structural analysis computer programs. A
simple procedure suitable for application in cases where only a small num-
ber of variables participate is described and demonstrated.
Purchased from American Institute of Aeronautics and Astronautics

Chapter 13

Stochastic Crack Growth Models


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I. Introduction

E XPERIENCE has shown that structures that are subjected to periodi-


cal loads fail after some period (or load cycles) at a stress that is
significantly lower than their static failure load. The phenomenon is known
as fatigue. It is common practice for designers to work with S-N curves,
which are empirical data relating the failure stress levels to the number of
loading cycles. The S-N curves are obtained by performing tests in which
the specimens are loaded by periodic (mostly harmonic) loads with levels
that are changed between specimens and observing the number of cycles
to failure. Many specimens are required for the generation of a good S-N
curve, and such curves can be found in many design manuals. The S-N
curves are, in fact, a collection of data points with a relatively high dispersion
that present a certain trend. When plotted on logarithmic scale axes, these
points usually present a trend of a decreasing straight line.
Practical designers can use S-N curves only as guidelines. The reason is
that usually the designed structure does not have geometrical and loading
conditions similar to the test specimens. Usually practical loadings are
not harmonic and certainly do not have constant amplitudes. A practical
structural element experiences wide spectrum loads, with a wide range of
frequencies and frequency-dependent PSDs. The oscillating load does not
always have a zero mean, and the ratio between maximum and minimum
amplitudes does not always coincide with fatigue tested specimens with
data contained in the manuals. During the past 30 years, theories for damage
accumulation were developed and applied. Most of these applications were
based on experimental observations, and during the early periods of the
design-to-fatigue most procedures were not explained theoretically.
Also during this period, models that try to explain the fatigue phenomena
by fracture mechanics were developed, e.g., Refs. 53-57. According to
these models, a very small crack that exists in the structure propagates
during each loading cycle, due to the high stresses that exist at its tip, which
are much higher than the average stress in the surrounding material. The
crack grows until its length is such that the complete structure fails. Thus,
the crack growth rate becomes higher toward failure, and at failure this
rate is infinite.
Models for crack growth have been the subject of thousands of papers
published over the past 30 years. These range from simplified models (such
Purchased from American Institute of Aeronautics and Astronautics

196 G. MAYMON

as in Ref. 58) to more advanced models (such as in Refs. 59 and 60). Some
of the models are based on the microstructure, and others are based on
the analysis of experimental data. Still others are based on the theory of
elasticity and use available experimental data for verification. Naturally, it
is beyond the scope of this book to quote the vast number of available
studies. In Ref. 61, an excellent list of 217 references is provided.
The experimental work described in many papers shows that the size of
a crack developed under repeated loads is of a random nature, although
a repeated trend is observed. This random nature is apparent even for the
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

experiments carried out under a strictly controlled laboratory setting using


deterministic external loads.
Four famous series of experimental test results are repeatedly quoted in
the literature: Noronha et a1.,62 in which two sets of experiments are pre-
sented; Virkler et and Ghonem and D ~ r eFrom . ~ ~the nature of the
results quoted in these references and many others, it can be seen that a
good prediction of the behavior of the crack growth process must be given
by a stochastic rather than a deterministic model.
In engineering practice, crack growth is usually treated by the determinis-
tic models. Expressions for the stress intensity factor required for such
analyses can be found in manuals, handbooks (such as Ref. 65), and text-
books (such as Ref. 66). Stochastic crack growth models have not been yet
adopted by the industry, although they describe the phenomenon better.
It is the purpose of this chapter to introduce the readers to some of the
concepts and procedures of the stochastic models, to show how these models
can be incorporated in the calculation of the probability of structural failure
due to crack propagation, and to encourage the engineering community to
study and adopt such methods in practical design.

11.Stochastic Crack Growth Models


There are three major reasons for the indeterministic nature of crack
growth process:
1) The macroproperties of different structures and specimens, that is,
the geometry, dimensions, and material properties (such as elastic moduli
and allowable stresses), may differ slightly between specimens. Thus, the
whole structure is indeterministic.
2) The external loading in practical engineering cases is usually random.
3) The microproperties of a structure or a specimen are random, which
means that the microstructure is not homogenous, even for strictly con-
trolled material production conditions.
This chapter focuses mainly on the third issue, although some aspects of
the first one are considered. The randomness of the external loads are not
treated here but are demonstrated in the last section of this chapter.
The measurement of crack size as a function of loading cycles (or time
as an interchangeable parameter) has the general form shown in Fig. 13.1.
The curves are characterized by two features: 1) a nonlinear increase in
crack size with an increased number of loading cycles (or time) and
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS 197


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I
Time or Load Cycles
Fig. 13.1 Generic curves for crack size as a function of time or cycles.

2) intermingling of the curves for several identical specimens. Curves similar


to those of Fig. 13.1 can be seen in Refs. 62-64.
Stochastic models for crack growth were suggested by many investigators
over the past 15 years, (e.g., in Refs. 67-73). These include evolutionary
probabilistic models (Markov chain, Markov diffusion models), cumulative
jump models (Poisson process, birth process, jump-correlated models), and
differential equation models. A comprehensive summary of the state of
the art is given in Ref. 61.
The models that can be used relatively easily in practical engineering
applications are the differential equation (DE) models. Two major groups
of models are used to treat the DE for the behavior of a stochastic structure
due to crack growth and to evaluate the probability of failure of the
structure:
1) Use the deterministic DE for the crack growth rate, while assuming
that the different parameters in this equation are random variables (RV).
Examples for this family are presented in Refs. 25, 74, and 75.
2) Use the modified DE for the crack growth rate, where a stochastic
nature of this rate is expressed by a random process (RP). Examples for
this family are presented in Refs. 66-70, 72, and 75-78.
For RV methods, the basic DE for the crack growth rate is
Purchased from American Institute of Aeronautics and Astronautics

G.MAYMON

where:

a = crack length
t = time
N = cycles of load
p = crack growth parameters
AKI = range of stress intensity factor
b = geometry parameters
q = parameters of the external load
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

The classic basic model of Ref. 58 is a special case of Eq. (13.la)

where Y ( a ) is a geometry function that depends on the mutual geometry


of the crack and a plate structure. D and n are considered as material param-
eters.
This model was used in Ref. 25 for the experimental results of Ref. 63,
in Ref. 67 for the results of Refs. 62 and 63, and in Ref. 61 for the results
of Refs. 63 and 64. When test results (crack size as a function of time or
cycles) are processed to yield crack growth rate as a function of the crack
length and plotted on logarithmic axes, experimental results are similar to
those shown in Fig. 13.2.

Time or Load Cycles (log scale)


Fig. 13.2 Generic curve for crack growth rate as a function of time or cycles,
Eq. (13.2).
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS 199


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Time or Load Cycles (log scale)


Fig. 13.3 Generic curve for crack growth rate as a function of time or cycles,
Eq. (13.3).

Another example of a D E for the crack growth process is given in Ref. 71


for the experimental results of Ref. 63,

where Ci, i = 1, 2, 3,4, are constants determined by the experimental data


shown generically in Fig. 13.3. The DE (13.3) extends the range of Eq. (13.2)
for values of a beyond the straight line region shown in Fig. 13.2. There
are many other examples of crack growth rate DE, e g , in Refs. 61 and
71, but they are not repeated here.
The solution for a stochastic crack growth rate using the RV methods
is performed by assuming a certain probability distribution for n and D in
Eq. (13.2). In Ref. 25 this was done for a center cracked plate, where n
was assumed to have normal distribution, D was assumed to have log-
normal distribution, and n and D were assumed to be correlated. In Ref.
79, a similar randomization was done for a notched beam loaded by a
bending moment. The disadvantage of this type of randomization is that
it does not simulate intermingling of the a-N curves, such as those shown
in Fig. 13.1.
Another disadvantage of the RV methods is that there is a physical
inconsistency when the constants n and D in the crack growth rate equation
(13.2) are randomized. The units of the stress intensity factor AKI are of
stress times square root of length, i.e., psi. in."2 or MPa. mU2.The crack
growth rate has units of length per unit time or per cycle. Thus, the units of
D in Eq. (13.2) are [length(1+3n'2)l(s~
forcen)] or [length('+3n'2)l(cycle.forcen)].
Purchased from American Institute of Aeronautics and Astronautics

200 G.MAYMON

If n is assumed random between specimens, the units of D become random.


This result has no physical meaning. The use of Eq. (13.3) presents similar
problems because sinh( ) and log( ) have a physical meaning only if their
argument is dimensionless; thus, randomization of the constants Cihas no
physical meaning. This difficulty can be overcome if, for instance, Eq. (13.2)
is written as a nondimensional equation, e.g.,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where K is a normalizing factor (which can be random) with dimensions


of the stress intensity factor, e.g., Young modulus multiplied by the square
root of a typical length. In this case, reported experimental results should
refer to this normalization factor and its probability density, when it is not
deterministic. At present, reported results do not include such information
and, therefore, the use of RV methods in engineering applications can yield
suspicious results.
An interesting almost nondimensional crack growth model is presented
in Ref. 80, the user manual of the NASAIFLAGRO computer program.
In Ref. 80, curve fitting for a very large number of materials is presented,
as well as crack models for many geometries. All of the data is included
in the database of the FLAGRO program, which enables an interactive
calculation of crack growth process, as well as boundary element analysis
of various geometries.
In the RP approach, the randomness of the crack growth is described
by adding a nonnegative RP X(t) to Eq. (13.la). Thus, Eq. (13.la) is
written as

where KI is the stress intensity factor, AKz the stress intensity factor range,
S the stress amplitude, R the stress ratio, and a the crack length.
Equation (13.2a) is written as

where Q includes Y(a) and AS, whereas b is related to n. It is then assumed


that Q and b are constants, determined from the results as described in
Fig. 13.2 by linear regression, and the dispersion around the straight line
(in logarithmic scales) is contributed only by the RP X(t), e.g., Ref: 81. To
achieve mathematical simplicity, X(t) is modeled as a positive lognormal
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS 201

RP, which takes values around unity. Taking the logarithm of both sides
of Eq. (13.6) the following is obtained:

=b . log(a) + log(Q) + log(X(t)) (13.7a)


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where

Y = log (daldt)
U = log(a(t))
s =lode)
z = log(X(t))
Except for the Z term in Eq. (13.7b), this equation describes a straight line
with slope b and a vertical axis intersection q, as shown in Fig. 13.2. The
Z = log X term describes the dispersion of data points around the straight
line. Z = log X is a normal RP with zero mean and standard deviation a,.
The constants b and Q can be estimated by linear regression of plots similar
to those in Fig. 13.2. The regression procedure also yields a value for
Once qog(,(,))
uz = qog(x(r)). is known, the mean and the standard deviation
of X(t) can then be calculated using the normal-to-lognormal conversion
formulas

In Ref. 67, test results that appear in Ref. 62 for aluminum fastener hole
specimens were analyzed and the following results were obtained:

It is interesting to reproduce artificially crack growth rate results by using the


results given in Eq. (13.9). To do this, the following procedure is performed:
1) RPs Z(a) with a normal distribution with ulogx = 0.087635 are created
Purchased from American Institute of Aeronautics and Astronautics

202 G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

a (inches)
Fig. 13.4 Four normally distributed RPs with a,,,, = 0.087635.

using a random generator computer program. For clarity, only four of these
types of processes are described in Fig. 13.4.
2) Using the equation (daldt) = Q X ab X l o Z [see Eqs. (13.6) and
(13.7)], values of the crack growth rate are computed. Plots of the logarithms
of these values as a function of a for four processes are described in Fig. 13.5.
3) A simple integration scheme in the a direction, beginning at an initial
crack length of a. = 0.004 in. is performed. Curves of a as a function of t
for eight cases are plotted in Fig. 13.6. In Fig. 13.7, an insert of Fig. 13.6
is plotted, which better shows the intermingling of the curves.
Using expressions of Ref. 81, it is also possible to prepare curves of
statistical limits to the possible a-t curves that have the statistical properties
as described in Eq. (13.9). In Fig. 13.8 such limits are described for the
problem in question. The curves dictate that the probability that a specimen
has a crack size growing faster than, for example, the line denoted lo%, is
10% and less.

111. Crack Length Distribution


To calculate probabilities of failure of a cracked structural element, the
distribution of the crack length is required. A very convenient model for
crack length distribution was presented in Ref. 67. The crack growth is
approximated by a Markov chain. Whereas this approximation introduces
an error associated with the crack growth rate, the error is negligible except
for very short fatigue life, which is of less interest for practical good designs.
These errors can be overcome if the model described in Ref. 70 is used.
Nevertheless, the simplicity of application of the model presented in Ref. 67
has many advantages when practical engineering cases are solved.
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS 203


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

-6.4
-6.6 1 ...................................................

I 1 I

-3 -2.5 -2 -1.5 -1
log(a)
Fig. 13.5 Crack growth rate as a function of crack length.

0 5000 loo00 15000 20000


Time (Flight Hours)
Fig. 13.6 Eight simulated crack growth curves.
Purchased from American Institute of Aeronautics and Astronautics

204 G.MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

3500 4000 4500 5000 5500 6000 6500


Time (Flight Hours)
Fig. 13.7 Insert from Fig. 13.6.

0 I I
--.- ----.+---
0 2500 5000 7500 10000 12500 15000 17500
Time (Flight Hours)
Fig. 13.8 Limits to crack growth curves with statistical data of Eq. (13.9).
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS 205

There are two extreme cases of the RP X(t) to be studied. When this
process is assumed to be completely independent at two moments in time,
it was shown that the statistical variability of the crack size is the smallest
possible. As a result, unconservative results that may not be used in engi-
neering applications are obtained. On the other hand, if the process is
assumed to be totally correlated at any two moments in time, this process
becomes an RV, and the statistical dispersion is the largest possible. The
reality is somewhere between these extreme cases, and some correlation
should be assumed.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

It is reasonable to assume that crack growth process is local and, there-


fore, is correlated only within a certain distance near the crack tip in the
structure. In Ref. 67 it was assumed that the random process X(t) can be
described by

where p is the mean of X(t) and Y(t) is an RP with zero mean. It was
assumed that the correlation coefficient of Y(t) has a triangular shape so
that there is no correlation beyond a time lag A. Thus, the coefficient of
correlation C, is described by

= 0; otherwise

The numerical value of A is selected in Ref. 67 so as to best fit given


experimental results. In other works, e g , Refs. 82 and 83, it is assumed
that the correlation coefficient is inversely proportional to a typical value
of the crack growth rate (thus, when the rate is large the correlation time
is small) and the correlation function has an exponential decay

cC= exp (- 6) iC = aoE($)


where Cc is the coefficient of correlation, E(da1dt) is a typical value of
daldt, and a. is a proportional constant. It was suggested to use a o =
0.32 mm = 0.012 in. for high tensile strength steel. In Fig. 13.9, Eq. (13.11)
is described for A = 8000, and Eq. (13.12) is shown for three values of
E(da1dt). Here log E(daldt) is selected as (-5.2), (-5.9, and (-5.8). These
are typical values (as can be seen from Fig. 13.5), which correspond to
T, = 2000, 4000, and 7950 flight hours, respectively.
Purchased from American Institute of Aeronautics and Astronautics

206 G.MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

0 2000 4000 6000 8000 10000 12000 14000


Correlation Time (Flight Hours)
Fig. 13.9 Correlation functions for several cases.

In Ref. 67, a DE for the PDF of a crack length a at time t is evaluated


and solved given that the crack length was a. at time to. This yields

X exp
2aZ A(t - to) '3
If the function a ( a ) is known and p, A, and a: are given, Eq. (13.13) can
be solved analytically or numerically.
According to the assumptions made in Ref. 67 for the evaluation of the
solution (13.13),values of a(t)smaller than a. can be obtained, an impossible
physical result, which means that the crack length may decrease. Limiting
the solutions to values a > a. yields a CDF that is larger than one. The
error is large for smaller values of t - to. If long life periods, which are of
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS 207

more interest in practical engineering work, are examined, the error is


small. Therefore, in Ref. 67 it is suggested to normalize Eq. (13.13) so that
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where

When g = Q . a b [see Eq. (13.6)],the integral in Eq. (13.14) can be solved


analytically. If the crack length at to is ao, the probability distribution of
the crack length q, at time t is given by

[(I ( a l a o ) - ( b - l ) l c ~ a-~ -' t (


- -
20: A(t - to)
-
forb +1 (13.15a)

D(t - to)
qa(a,tlao, to) =
Q a b d 2 n a ; A(t - to)

where c = b - 1. Q and b are the coefficients appearing in Eq. (13.6), A


is the correlation time of the process, and 0, is the standard deviation of
the process X(t) [see Eq. (13.8)].
In Fig. 13.10, the PDF of the crack length is shown for different times,
with A = 8000 flight hours, a value selected in Ref. 67 to best fit experimental
results of Ref. 62.
The CDF of the crack length can be determined using

Numerical integration of Eq. (13.16) was used to obtain the CDF shown
in Fig. 13.11. When a crack length is selected, the probability that the crack
Purchased from American Institute of Aeronautics and Astronautics

208 G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Crack Length (inches)


Fig. 13.10 PDFs of crack length.

Crack Length (inches)


Fig. 13.11 CDF of crack length, for different flight hours.
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS 209


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Crack Length (inches)


Fig. 13.12 Probability of a crack being higher than a given value, for different
flight hours.

length is lower or equal to this value for a given time can be determined
from Fig. 13.11.
The probability that a crack is larger than a given value is

These probabilities are shown in Fig. 13.12.


In the described model, a relatively simple expression (13.15) is presented
for a crack that follows a crack growth rate equation such as Eq. (13.6).
This expression can be used in many practical applications and is demon-
strated later in this chapter. More advanced crack models can be found in
the literature, e.g., Refs. 82 and 83. These models assume randomness of
the initial crack size and a threshold value of initial crack size below which
no crack growth takes place. Also, interaction between the RP of the crack
propagation and the random loading process may be taken into account.
Although these models provide a more complete treatment of the crack
length PDF, they are much more difficult to apply when an analysis of the
probability of failure of a practical design is evaluated. The simplicity of
Eq. (13.6) enables its application in the probabilistic analysis programs and,
thus, a practical analysis can be performed. The reader who is interested
in more advanced models is refered to the list of references, especially to
Ref. 61.

IV. Failure Criterion


The selection of a failure criterion for a certain analyzed structure is
one of the most important decisions in engineering practice. This decision
Purchased from American Institute of Aeronautics and Astronautics

210 G. MAYMON

influences the nature of the design, the amount of marginal safety, and the
prediction of the structure's behavior. It is this decision that requires the
talent, expertise, intuition, and experience of the design engineer. In many
designs it is common to require that the maximum stress in the structure
will never exceed the material's yield stress (with a given margin of safety
or with a predefined probability of failure). As crack growth is a process
in which local plastic zones exist and local yielding of the structural element
may be permitted, this is not a well-suited criterion for the phenomenon.
A very simple criterion can be expressed if the demand is
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

where Klc is the critical stress intensity factor and KI is the stress intensity
factor. The critical stress intensity factor is the stress intensity factor calcu-
lated when the crack size is such that an infinite rate of crack growth exists
and the structure is ruptured. If the stress intensity factor, for a given
structural element geometry, is known as a function of the crack length,
the critical stress intensity factor can be calculated.
The failure function g is then

where failure takes place when g 5 0. KI is a function of the structure


characteristics and the loading, e.g., Eq. (13.2b). Both KIc and KI can be
random. This failure criteria is now required by aerospace regulations.
A more advanced criterion combines the brittle fracture, ductile fracture,
or mixed failure of metals, which was presented in Ref. 84, the Burdekin-
Stone criterion. This provides the following relation for the safety and
failure regions:

where a, is the yield stress, u,is the ultimate tensile stress, and KIc is the
critical stress intensity factor, for which expressions are provided in the
following equations. KL corresponds to linear elastic failure and SP corres-
ponds to plastic failure. Equation (13.20) is described in Fig. 13.13.
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS 21 1

- SAFE

-
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

-
I
I
I
I
I
I
I
I
1
I
,
I
I I
I
I

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

P
Fig. 13.13 Safety and failure regions (Burdekin-Stone criterion).

In Ref. 85, expressions for KIc are given for a centered cracked plate
(initial crack length 2ao, plate width b) under uniform tensile stress a

where

[ 42J
(:
ub,= a, 1 - - - a,,
I (zone 1)

(zone 2)

, = K & 1 - ) (zone 3)

KC1is the fracture toughness, which is a material property and can be


found in material manuals. For demonstration, Eq. (13.22) is described in
Fig. 13.14 for typical values: a, = 75,400 psi, Kc* = 100,700 lb/in.1.5, and
b = 12 in.
To demonstrate the calculation of the probability of failure of a structural
element, the failure criterion (13.20) is used in Sec. V. Use of Eq. (13.19)
yields results that are more conservative.
Purchased from American Institute of Aeronautics and Astronautics

212 G. MAYMON

80000 ,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

0 1 2 3 4 5 6
Initial Crack Length (inches)
Fig. 13.14 Stress ua,as a function of initial crack length.

V. Probability of Failure Calculation


The quantities and distributions defined and calculated in Sec. IV and
the procedures described in the preceding chapters are used here to calcu-
late the probability of failure of a centered cracked plate.
A plate of width b is subjected to a tensile stress crand has a centered crack
of length 2ao. It is assumed that the crack is in zone 1. Using Eq. (13.20) a
failure function g can be written as

Failure occurs when g 5 0. Using the classical expression for a centered


cracked plate, with a very small crack compared to the width b, the geometry
function Y ( a ) can be assumed to be a unity; thus,
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS 213

Introducing Eqs. (13.20-13.22) and (13.24) into Eq. (13.23) yields


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Denoting

the failure function can be written as

The calculation is made with the crack length PDF described in Fig. 13.10,
for 4000 flight hours. It should be noted that this curve was obtained by
processing the experimental results of an aluminum fastener hole62and not
of a centered cracked plane and, therefore, the curve is used here for
demonstration purposes only. The relevant distribution curve of Fig. 13.10
is not included in the distribution library of the probabilistic structural
analysis programs and, therefore, to use these libraries, the curve was
approximated by a lognormal PDF. In Fig. 13.15 the approximation is
shown together with the original curve.
The initial crack size is assumed as deterministic, a. = 0.004 in. Table 13.1
provides data for the distribution and their first two statistical moments of

Table 13.1 Data for stochastic crack growth example

Units Distribution Mean Standard deviation

XI psi Normal Varies Varies


x2 psi Weibull 75,400 3016
x
3 psi Weibull 97,157 3886
x
4 Iblin.' Weibull 100,700 10,070
x5 in. Deterministic 0.004 n.a.
X6 in. Lognormal 0.00805 0.002
Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

0 0.005 0.01 0.0 15 0.02


Crack Length (inches)
Fig. 13.15 Lognormal approximation (+ + + +) to PDF (-) of Fig. 13.10.

Mean Value of External Load (psi)


Fig. 13.16 Probability of failure after 4000 flight hours; cov = 0.05: lower line
FORM, upper line SORM, (+) linear simulation, and (0)nonlinear simulation.
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS 215


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

COV of Loading (%)


Fig. 13.17 Probability of failure after 4000 flight hours, XI = 35,000 psi: lower line
FORM, and upper line SORM.

the random variables. Equation (13.27) with the data of Table 13.1 is
introduced into the CALREL program,37 and the probabilities of failure
after 4000 flight hours are calculated.
First, the standard deviation of XI is assumed to be 5% of the mean
(cov = 0.05). The probability of failure after 4000 flight hours is shown in
Fig. 13.16, as calculated by FORM, SORM, and linear and nonlinear direc-
tional simulation (20,000 simulations). Then, a mean values of 35,000 psi
is assumed for X I , and different values are assumed for the coefficient of
variation. Results for this calculation are shown in Fig. 13.17.
The small difference between the FORM and SORM calculations means
that despite the nonlinearity of the failure function (13.27), the calculated
results are not influenced by this nonlinearity. Also, good agreement is
obtained between simulation and approximate methods.

VI. Required Information for the Analysis


In the preceeding paragraphs, a stochastic analysis of a cracked structural
element was demonstrated, and the probability of failure was calculated.
This process used some important items of information that are required
to apply the described procedure. These items are as follows.
1) An expression, or rather a method of calculation, of the stress intensity
factor, e.g., (Eq. 13.2b), is necessary. For many geometries such expressions
can be found in Ref. 65. Also, many finite element computer codes can
compute this factor numerically. Then an approximate closed-form expres-
Purchased from American Institute of Aeronautics and Astronautics

216 G. MAYMON

Define or select the DE for crack growth rate,


Sec. I1
1

process X(t), Sec. I1


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Built a PDF for the crack length for different times,


Sec. I11

I Select an appropriate failure criterion, Sec. IV


I
Build a failure function, Secs. IV and V
I

*I
I Calculate the probability of failure, Sec. V
I
Fig. 13.18 Process for probabilistic calculation of cracked structural element.

sion can be determined using a Taylor expansion or any other curve fit-
ting method.
2) A crack growth DE is necessary. Equations (13.1) are of general form.
Equation (13.2b) is only a simple example, used for demonstration purposes
in this presentation.
3) An estimation of the correlation time of the stochastic process X ( t )
is necessary to define, for instance, the parameter A of Eq. (13.13). In Ref.
67, a triangle function was assumed, leading to Eq. (13.13), and the proper
A was selected so as to best fit experimental results. In Ref. 82 it was
assumed that the correlation time is inversely proportional to a typical
value of the crack propagation rate, and the correlation function is an
exponential decay. Thus, an estimation of r, in Eq. (13.12) is required for
a solution. These required parameters must be estimated from experimental
data of the particular case under consideration, or a similar known case.
4) It is important to establish the failure criterion. An example was
presented in Sec. IV, but many others can be thought of. As was already
Purchased from American Institute of Aeronautics and Astronautics

STOCHASTIC CRACK GROWTH MODELS

mentioned, the determination of the failure criterion is of crucial impor-


tance.
5 ) Statistical data on the yield and tensile strength, fracture toughness,
and initial cracks must be made available. Some constants, such as Q and
b in Eq. (13.6) and the value of qogx(r) were determined using experimental
data. In cases where experimental data is not available, these values can
be estimated based on reported results of similar materials.
At present it would be realistic to limit the demonstrated procedure to
metal and metallike materials. This type of analysis is much more compli-
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

cated if a composite structure is to be analyzed. The mechanism of crack


growth cannot be described by a simple DE, and the calculation of the
stress intensity factor is much more complex. Little experimental data is
available in the literature. Also in the present analysis, effects of spectral
load amplitudes on crack growth response, such as acceleration and crack
delay, are not included. Thus, the method presented is applicable to load
spectra that are stationary, where time histories of the load do not include
sudden jumps.
In Fig. 13.18 a general flow chart that demonstrates the required analysis
steps, as well as the related paragraphs in this chapter, is shown. The flow
chart is applicable to the general case.

VII. Summary
A procedure for the calculation of the probability of failure of a cracked
structural element is described and demonstrated.
Although the principles of the solution process are general, they were
demonstrated only for metal and metallike materials.
Random process methods are recommended for the stochastic analysis
of crack growth. .
A relatively simple model for the calculation of the PDF of crack length
is described. There are many other models that may describe the phenom-
ena more rigorously, but they are too complex for daily routine applications.
The demonstrated process is limited to stationary loadings.
Some important required information is listed.
It is important to bear in mind that many unknown factors still exist in
this field, and results should be treated carefully. This type of analysis is
suited especially for comparison between several design options.
It is very important to form a data bank for the statistical parameters
required in this type of analysis.
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

Chapter 14

Concluding Remarks
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

A N effort was made in this book to introduce readers to the basic


concepts of the analysis of random vibration and random structures.
The author tried to direct readers to practical use of the fundamental
concepts for the daily engineering work. The book is written for design
engineers who have a basic understanding of random vibration theory, and
it demonstrates the practical application of the material discussed in each
section. Readers are introduced to a variety of solution techniques and are
directed in how to formulate their own problems, how to solve them, and
what is the methodology of the solution.
One of the major purposes of the presented material is to introduce
readers to the physical meaning of the theoretical models of random vibra-
tion and random structures. Many (but not all) of the subjects described
in the preceding chapters can be solved today by commercially available
computer programs; some of these are mentioned in the text. These pro-
grams must be considered as part of the tools used by the design engineer
in the process of the design. As a complete numerical analysis of every
alternative and every possible option of a practical design is prohibitive,
the understanding of the meaning of the structure's behavior and the basic
assumptions of the theories (on which the numerical analysis is also based)
may help the engineer to formulate simple models. With such models the
effects of various design parameters and the sensitivity of the design to
changes in these parameters can be performed very quickly, especially in
the preliminary design stage. Heavy numerical analysis should be kept for
verification of the close-to-final design, where detailed refinements (which
are more difficult to be formulated with simple expressions) can be fi-
nally checked.
The author hopes that the material presented will encourage young
engineers to explore the interesting fields of random vibration, random
structures, and structural probabilistic analysis. These fields are of major
importance in the design of any structural system, and it is believed that
their importance will grow in the near future.
The understanding of the basic concepts and their physical meaning will
undoubtedly result in better designs and, therefore, with better designers.
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

Appendix A

Some Important Integrals


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

T HE integrals summarized in this appendix are sometimes required in


the solution of the response of MDOF and continuous systems to white
noise, one-sided white noise, band-limited white noise, one-sided band-
limited white noise, and filtered white noise, described in Chapter 4, Sec.
111, and for the calculation of spectral moments described in Chapter 12,
Sec. 11. These integrals are based on Der Kiureghian.13
I. Constant Value PSD Functions
For the direct terms

n 1 - (2127) arctan (&I-)

For mixed (interaction) terms

= j,"o ~ e [ g ~ ( w ) E ? ( odo])
Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON

In Eqs. (A4-A6),

Equations (A4-A6) degenerate to Eqs. (Al-A3), respectively, when i = j.

11. Filtered One-sided White Noise


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

For this section, see Eq. (4.22).


For the direct terms

where, for r = milog.


Purchased from American Institute of Aeronautics and Astronautics

APPENDIX A

For mixed (interaction) terms


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

+ C1[2alblA3 (a?- b?)A4]+ C2[2alblA4+ (a: - b:)A3]


-

blD2

In these equations

A, = [(a,- al)' - (bg - b:)][(ao+ a2)' - (bi - bz)]- 4bz(a0- al)(ao+ a2)

A2 = 2bo(ao+ a2)[(a0- a,)' - (b; - b?)]+ 2bo(ao - al)[(ao+ a2)' - (bi - b;)]
Purchased from American Institute of Aeronautics and Astronautics

224 G . MAYMON

A3 = [(al + ao)' - (b: - b$)][(al+ a2)' - (b: - b;)] - 4b?(a0 + al)(al + a2)

A4 = 2bl(al + a2)[(al+ ao)' - (b: - b;)] + 2bl(al + ao)[(al + a2)' - (b? - b;)]

A, = [(a, + ao)' - (b; - bT)][(ao- a2)' - (b$ - b?)] - 4b$(a0+ al)(ao - a2)

A6 = 2bo(ao- a2)[(ao+ al)' - (b; - b:)] + 2bo(ao+ al)[(ao- a2)' - (b; - b;)]
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

A7 = [(al - aol2- (b: - b;)][(al+ ao)' - (by - b:)] - 4b:(a0 + al)(al - ao)

As = 2bl(al + ao)[(ao- al)' - (b? - b;)] + 2bl(al - ao)[(ao+ all2 - (b: - b$)]

A, = [(a2- al)' - (b; - b;)][(a2+ ao)' - (b; - b;)] - 4bg(a2 - ao)(a2+ ao)

A,, = 2b2(ao + a2)[(a2- aO)' - (b? - b;)] + 2b2(a2- ao)[(a2+ ao)' - (b; - b;)]
Purchased from American Institute of Aeronautics and Astronautics

APPENDIX A

These expressions can be carefully programmed into a very simple com-


puter program.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

Appendix B

Conversion Between Acoustic Decibels and PSDS6


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

T RADITIONALLY, SPL is described by the root mean square levels


of the pressure fluctuation as a function of the frequency. The units
of these levels are acoustic decibels. The definition of acoustic decibel is

PI,,
acoustic decibel = 20 log -
Pref

where PI,, is the root mean square of the pressure fluctuations, and Pref=
2 x dyne/cm2 = 2 X Pa = 2.0408 X kg/m2 = 2.0408 X
lo-'' kg/cm2.
Acoustic SPLs are measured by special instruments in which the fre-
quency is scanned with a certain bandwidth. Two kinds of bandwidths are
usually used, octave and 113 octave. An octave is a bandwidth in which
the upper frequency is twice the lower one. A 113 octave is one-third of
the bandwidth of an octave, where the range between the upper and lower
frequencies are divided into three bands, equally spaced on a logarithmic
scale.
Because PSD is squared pressure per hertz and the measurements are
done for a finite bandwidth Af, which is different from 1 Hz, a correction
term must be introduced when converting from acoustic decibel to
pressure2/hertz and vice versa. The following equations apply:s6
LPS = SPL - AL (B2)
where SPL is given (measured) and A L is a correction term

where Af is the bandwidth (octave or 113 octave).


By the definition (Bl)

PI,,
LPS = 20 log -
Pref

from which P,,, can be obtained. This PI,, , when squared, yields the PSD
of the excitation
PSD = (Pr,,)2/(1 Hz) (B5)
Purchased from American Institute of Aeronautics and Astronautics

228 G. MAYMON

I External data I
Graph of sound pressure level New data
as a function of frequency in
Hz, 1/3octave fi,Afi,SPLi

+
Divide the frequency range fmd A Li using (B.3)
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

into 113 octave bandwidth


strips, A f i
find LPSi using (B.2)

For each bandwidth, find the


center frequency f
I f i d Prm% using (B.4) I
For each f find the
corresponding SPLi from the
external data
+
a PSDi(f)=(Prmsi)*
can be converted to
PSD,(o)

Fig. B1 Conversion scheme (acoustic decibels to PSD).

The PSD is obtained in pressure2/hertz as a function of the frequency f (in


hertz). This should be converted to pressure2/(radian/second) by dividing
the values by 227, and changed to angular frequency o by multiplying the
frequency (hertz) by 227.
Figure B1 schematically described this procedure. In cases where PSD
is to be converted into acoustic decibel, the same equations can be used.
Purchased from American Institute of Aeronautics and Astronautics

Appendix C

Finite Element Input Files for MJPDF Method


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

I N THIS appendix, the input files used in the examples described in


Chapter 12 are listed. The examples were computed on the ANSYS 4.2
finite element program. Inasmuch as the present version of ANSYS may
be different (in 1996, Version 5.3 was already available) the following files
must be properly modified according to the user's available version. When
another finite element program is used, the following file may be used only
as examples and guidelines for the preparation of compatible files. In several
lines comments are added to improve understanding of the file.
I. File for Example 1
BEGIN-INP
ml=1000
sl=33
m2=2
data from Table 11.1
s2=0.1
mz=600
sz=30
deml = (6.283185307)*((~1)*(~2))
dem2=sqrt(6.283185307)
dem3 =(dem2)*(sz)
*set,x1,1049.5 initial value for F
*set,x2,1.85 initial value for A
/PREP7
ax1 =(-0.5)*(((~1-ml)*(x1-ml))/(sl*sl)) exponent of normal dist. of F
ax2= (-0.5)*(((~2-m2)*(~2-m2))/(s2*s2)) exponent of normal dist. of A
a=(axl) + (ax2)
y = (exp(a))/(deml)
et,l,l
ex,1,2.le6
r,l,x2
n,1,0,0 finite element modeling
n,2,10,0
e,L2
d,l ,all,O
f,2,fx,xl
afwrite
1 constraint at bar end
axial force on the other end

finish
Purchased from American Institute of Aeronautics and Astronautics

230 G. MAYMON

/INPUT,27 solution phase


finish
/POST1 postprocessor
stress,axst,l,3 axst is axial stress
set,l
*get,z,axst,l z is axial stress at node 1
az= (-OS)*(((z-mz)*(z-mz))/(sz*sz))
zz= (exp(az))l(dem3) first part of Eq. (12.24)
t=(y)*(zz> MJPDF
tt= (l)/(t) inverse of MJPDF-the objec-
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

tive function
finish
/OPT optimization module
opvar,xl,dv,1000,1099,0.2 dv variable, range 1000-1099
opvar,x2,dv,1.7,2,0.001 dv variable, range 1.7-2
opvar,z,sv,540,600,1 sv variable, range 540-600
opvar,tt,obj,,,0.001
oplist
OPCoPY
opeqn,,,J
oprunP0
oplist,all
finish

11. File for Example 2


BEGIN-INP

1
m l = 1000
sl=33
m2=2

J
s2=0.1
m3=50 data, Table 11.3
s3=2
my =33000
sy= 1000
p1 =I574960995 (~qrt(257))~
p2=2.506628275 sqrt(2n)
*set,x1,104Y.5
*set,x2,1.85
*set,x3,53
/PREP7
] initial values for load effects
term

ax1=(-OS)*(((xl-ml)*(xl -ml))/(sl*sl)) exponent of normal dist. of F


ax2= (-0.5)*(((~2-rn2)*(~2-m2))/(~2*~2)) exponent of normal dist. of I
ax3= (-0S)*(((x3-m3)*(~3-m3))/(~3*~3)) exponent of normal dist. of L
deml =(pl*sl)*(s2*s3)
dem2= (p2)*(sy)
q =((axl) +(ax2)) + (ax3)
Y =(exp(q)/(deml) second part of Eq. (11.29)
Purchased from American Institute of Aeronautics and Astronautics

APPENDIX C

et,1,3
ex,1,2.le6
a=3*x2
r,l,a,x2,2
n,1,0,0
n,ll,x3,0
fill,l,ll
e,1,2
egen,lO,l,l
d,l,all,O
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

f,ll,fy,xl
afwrite
finish
/INPUT97 solution phase
finish
/POST1 postprocessor
stress,bend,3,10 bend is bending stress
set
*get,z,bend,l z is bending stress at node 1
ax=(-OS)*(((z-my)*(z-my))/(sy*sy))
zz= (exp(az))/(dem2) first part of Eq. (11.29)
t=(y>*(zz> MJPDF
tt =(l)/(t) inverse of MJPDF-the objec-
tive function
finish
/OPT optimization module
opvar,xl,dv,1000,1099,1 dv variable, range 1000-1099
opvar,x2,dv,1.7,2.0,0.001 dv variable, range 1.7-2
opvar,x3,dv,50,56,0.1 dv variable, range 50-56
opvar,z,sv,30000,36000,100 sv variable, range 30,000-36,000
opvar,tt,obj
oplist
OPCOPY
opeqn,,,,l
oprun,40
oplist,all
finish

111. File for Example 3


BEGIN-INP
ml=1000
sl=33
m2=2
s2=0.1
my=600
sy =20 A
r=0.5 correlation coefficient
r2=(r)*(r)
rd = (2) * (r)
p2= 6.283185307
Purchased from American Institute of Aeronautics and Astronautics

G.MAYMON

s=sqrt((l)-(r2))
p3 =2.506628275
ci=((l)-(r2))*(-2)
c=(l)l(ci)
*set,x1,1033
*set,x2,1.85
/PREP7
a1 =((xl -ml)*(xl-ml))/(sl*sl)
a3 = ((x2 -m2)*(x2 -m2))/(s2*s2)
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

a21 =((XI-ml)*(x2-m2))
a22=(sl)*(s2)
a2 = ((rd)*(a2l))/(a22)
p l = ((al) - (a2)) + (a3)
p=(c)*(pl)
deml = ((p2)*(sl)*((s2)*(s))
Y = (exp(p))/(deml) second part of Eq. (11.31)

1
et,l,l
ex,1,2.le6
r,l,x2
n,1,0,0
n,2,10,0
finite element modeling
e,1,2
d,l,all,O
f,2,fx,xl
afwrite
finish
/INPUT,27 solution phase
finish
/POST1 postprocessor
stress,axst,l,3 axst is axial stress
set
*get,z,axst,l z is axial stress at node 1
az= (-0.5)*(((z-my)*(z-my))/((sy)*(sy)))
dem2= (p3)*(sy)
zz= (exp(az))/(dem2) first part of Eq. (11.31)
t=(y)*(zz) MJPDF
tt = (l)/(t) inverse of MJPDF-the objec-
tive function
finish
/OPT optimization module
opvar,xl,dv,1000,1040,1 dv variable, range 1000-1040
opvar,x2,dv,1.7,.2.0,0.001 dv variable, range 1.7-2
opvar,z,sv,540,600,1 sv variable, range 540-600
opvar,tt,obj
oplist
OPCOPY
opeqn,,,,l
oprun,40
oplist,all
finish
Purchased from American Institute of Aeronautics and Astronautics

APPENDIX C

IV. File for Example 4


BEGIN-INP
p l = - 118.213806
p2 = 0.287027632
p3 = -0.00023850733
p4=6.978248213e-8
p5 =2.506628275
del=4.8587e-5
1
A
coefficients in Eq. (11.39b)

sqrt(2n)
coefficients in Eq. (11.38)
beta=2.1013491
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

gama=900 lower limit of Weibull dist.


ml = 1000
sl =50
m2=2
s2=0.1
my=700
sy=33
r0=0.408828
r2=0.817656
deml =OX32859666
dem2=5.73410688
*set,x1,1100
*set,x2,1.85

/PREP7
z2= ((x2) -(m2))/(s2)
zll=(pl)+((p2)*(xl))
zl2=(p3)*((xl)*(xl))
zl3=((p4)*(xl))*((xl)*(xl))
zl=((zll)+(zl2))+(z13) Eq. (11.39b)
a1 =(zl)*(zl)
a2= (z2)*(z2)
a12=((-l)*(r2))*((zl)*(z2))
aa=((al)+(a12))+(a2) terms in Eq. (11.41)
aaa=(-OS)*((aa)i(deml))
fzz= (exp(aaa))/(dem2)
ax2= (-0.5)*((~2-rn2)*(~2-m2))/(~2*(~2)
fx2= exp(ax2))/((~5)*(~2))
bl =(xl)-(gamma)
b2=(bl)**(beta)
b3=(-del)*(b2)
fl=exp(b3)
b4=(bl)**((beta)-(1))
fxl = ((del)*(beta))*((b4)*(fl))
fzl = (exp((-0.5*(al)))/(p5)
fz2= (exp((-OS)*(a2)))/(p5)

fT,* in Eq. (11.43)


Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON

et,l,l
ex,1,2.le6
r,l,x2
n,l,0,0
n,2,10,0
eS,2
d,l,all,O
f,2,fx,xl
afwrite
finish
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

4
/INPUT,27 solution phase
finish
/POST1
stress,axst,l,3 postprocessor
set
*get,z,axst,l
ax3= (-0.5)*(((z-my)*(z-my))l((sy*sy)))
fx3=(exp((ax3)>/((ps>*(sy)) first part of Eq. (11.43)
jj = (fx3)*(j) MJPDF
tt=(l)/(jj) inverse of MJPDF-the objec-
tive function
finish
/OPT optimization module
opvar,xl,dv,950,1200,0.1
opvar,x2,dv,1.7,2.0,005
opvar,z,sv,600,700,0.1
opvar,tt,obj
oplist
OPCOPY
opeqn,,,,l
oprun,40
oplist,all
finish

V. File for Example 5


BEGIN-INP
m l = 1000
sl=33
m2=2
s2=0.1
m3 =570
deml =(6.283l85307)*((sl)*(s2))
*set,x1,1049.5
*set,x2,1.85
IPREP7
axl=(-0.5)*(((xl -ml)*(xl -ml))/(sl*sl))
ax2=(-0.5)*(((x2-m2)*(x2-m2))l(s2*s2))
a = (axl)+ (ax2)
y =(exp(a))/(deml)
et,l,l
ex,1,2.le6
r,l,x2
Purchased from American Institute of Aeronautics and Astronautics

APPENDIX C

n,1,0,0
n,2,10,0
eJ,2
d,l ,a11,0
f,2,fx,xl
afwrite
finish
/INPUT97 solution phase
finish
/POST1 postprocessor
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

stress,axst,l,3
set
*get,z,axst,l
*set,yy,y
tt=(l)/(yy)
finish
/OPT optimization module
opvar,xl,dv,1000,1099,0.2
opvar,x2,1.7,2,0.001
opvar,z,sv,569.6,570,0.1
opvar,tt,obj,,,le-6
oplist
OPCOPY
opeqn,,,,l
oprun,40
oplist,all
finish

VI. File for Example 6


BEGIN-INP
p1 =l.l56517643
p2= 1.313035285
p3=0.850918128 coefficients in Eq. (11.51)
p4=8.199453e-11
p5=5.8042578e-10
p6=2.50662875 sqrt(2~)
p7=76.18028355 ( s q r t ( 2 ~ ) ) ~ x d e t [ REq.
]~.~
(11.50)

four parts of mean and standard


deviations of P

mean of El
standard deviation of EI
Purchased from American Institute of Aeronautics and Astronautics

G. MAYMON

creation of Eq. (11.51)


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

first part of Eq. (11.52)

second part of Eq. (11.52)


MJPDF
inverse of MJPDF-the objec-
tive function
et,1,3
ex,l,ee ' finite element modeling
r,1,1.333333,1.3
n,1,0,0
n,33,32,0
fi11,1,33
e,12
egen,32,1,I
d,l,all,O
d,33,a11,0
p,1,2,x1,,8
~,9,10,~2,,16
p,17,18,x3,,24
p,25,26,x4,,32
afwrite
finish

nNPUT,27
finish solution phase

postprocessor
mom is the bending moment
Purchased from American Institute of Aeronautics and Astronautics

APPENDIX C 237

set
*get,ss,mom,l ss is the bending moment at
node 1
finish
/OPT optimization module
opvar,xl,dv,8,15.2,0.001
opvar,x2,dv,8,15.2,0.001
opvar,x3,dv,8,15.2,0.001
opvar,x4,dv,8,15.2,0.001
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

opvar,ss,sv,1096,1100,1
opvar,tt,obj
oplist
OPCoPY
opeqn,,,S
oprun,40
oplist,all
finish
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

References

' Timoshenko, S., and Young, D. H., Vibration Problems in Engineering,


Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

3rd ed., D. Van Norstad, Princeton, NJ, 1955.


Den Hartog, J. P., Mechanical Vibrations, McGraw-Hill, New York,
1956.
Hurty, W. C., and Rubinstein, M. F., Dynamics of Structures, Prentice-
Hall, Englewood Cliffs, NJ, 1964.
Biggs, J. M., Introduction to Structural Dynamics, McGraw-Hill, New
York, 1964.
Thomson, W. T., Theory of Vibration and Applications, Prentice-Hall,
Englewood Cliffs, NJ, 1975.
Vierck, R. K., Vibration Analysis, Harper and Row, Addison, IL, 1979.
' De Vries, G., "Emploi de la methode vectorielle d'analyse dans les
essais de vibration," La Recherche Aeronautique, Vol. 74,1960, pp. 41-47.
Beatrix, C., "Les Procedes Exprimentaux de I'essai Global de Vibration
d'une structure," La Recherche Aeronautique, Vol. 109, 1965, pp. 57-64.
Maymon, G., "Response of Aeronautical Structures to Random Acous-
tic Excitation-a Stress Mode Approach," Random Vibration-Status and
Recent Developments, edited by I. Elishakoff and R. H. Lyon, The Stephen
Harry Crandall Festschrift, Studies in Applied Mechanics, Vol. 14, Elsevier,
New York, 1986, pp. 267-277.
lo Prazen, E., Modern Probability Theory and its Application, Wiley, New
York, 1960.
" Elishakoff, I., Probabilistic Methods in the Theory of Structures, Wiley,
New York, 1983.
l2 "Environmental Test Methods and Engineering Guidelines," MIL-
STD-810D, U.S. Department of Defense, 1983.
l3 Der Kiureghian, A., "Structural Response to Stationary Excitation,"
Journal of Engineering Mechanics, Vol. 106, No. 6, 1980, pp. 1195-1213.
l4 Crocker, M. J., "The Response of Supersonic Transport Fuselage to
Boundary Layer and to Reverberant Noise," Journal of Sound and Vibra-
tion, Vol. 9, No. 1, 1969, pp. 6-20.
l5 Maestrello, L., "Measurement and Analysis of the Response Field of
Turbulent Boundary Layer Excited Panels," Journal of Sound and Vibra-
tion, Vol. 2, No. 3, 1965, pp. 270-292.
l6 Schlichting, H., Boundary Layer Theory, McGraw-Hill, New York,
1955.
l7 Maymon, G., "Response of Plate-Like Structures to Correlated Ran-
dom Pressure Fluctuations," Proceedings of the 33rd AIAA/ASME/ASCE/
Purchased from American Institute of Aeronautics and Astronautics

240 G. MAYMON

A H S / A S C Structures, Structural Dynamics, and Materials Conference,


AIAA, Washington, DC, 1992, pp. 1741-1746.
ls Young, W. C., Roark Formulas for Stress and Strains, 6th ed., McGraw-
Hill, New York, 1989.
l9 Stoker, J. J., Nonlinear Vibrations, Interscience, New York, 1950.
20 Chu, H. N., and Hermann, G., "Influence of Large Amplitudes on the
Free Flexural Vibration of Rectangular Elastic Plates," Journal of Applied
Mechanics, Vol. 23, 1956, pp. 532-540.
21 Mei, C., "Response of Nonlinear Structural Panels Subjected to High
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Intensity Noise," US. Air Force Wright Lab., Technical Rep. AFWAL-
TR-80-3018, Wright-Patterson AFB, OH, 1980.
22 Maymon, G., and Rehfield, L. W., "Non-Linear Vibration of Clamped-
Clamped Beam," Proceedings of the 19th A Z A A / A S M E / A S C E / A H S / A S
Structures, Structural Dynamics, and Materials Conference, AIAA, New
York, 1978.
23 Maymon, G., and Rehfield, L. W., "Experimental Determination of
the Effect of Non Linear Stiffness on the Vibration of Elastic Structures,"
Israel Journal of Technology, Vol. 22, No. 1, 1984185, pp. 31-37.
24 Roberts, J. B., and Spanos, P. D., Random Vibration and Statistical
Linearization, Wiley, New York, 1990.
25 Madsen, H. O., Krenk, S., and Lind, N. C., Methods of Structural Safety,
Prentice-Hall, Englewood Cliffs, NJ, 1986.
26 Thoft-Christensen, P., and Baker, M., Structural Reliability Theory and
its Applications, Springer-Verlag, Berlin, 1982.
27 Melcher, R. E., Structural Reliability Analysis and Prediction, Ellis
Honvood, Chichester, England, UK, 1987.
28 Bjerager, P., "On Computational Methods for Structural Reliability
Analysis," New Directions in Structural Systems Reliability, edited by
D. Frangopol, Univ. of Colorado Press, Boulder, CO, 1988, pp. 52-67.
29 Shinozuka, M., "Basic Analysis of Structural Safety," Journal of Struc-
tural Engineering, Vol. 109, No. 3, 1983, pp. 721-740.
30 Cornell, C. A., "A Probability Based Structural Code," Journal of the
American Concrete Institute, Vol. 66, No. 12, 1969, pp. 974-985.
31 Hasofer, A. M., and Lind, N. C., "Exact and Invariant Second Moment
Code for Materials," Journal of Engineering Mechanics, Vol. 100, No. 1,
1974, pp. 111-121.
32 Rackwitz, R., and Fiessler, B., "Structural Reliability Under Combined
Random Load Sequences," Computers and Structures, Vol. 9, No. 5,
1978, pp. 489-494.
33 Chen, X., and Lind, N. C., "Fast Probability Integration by Three
Parameters Normal Tail Approximation," Structural Safety, Vol. 1, 1983,
pp. 269-276.
34 Hoenbichler, M., and Rackwitz, R., "Non Normal Dependent Vectors
in Structural Safety," Journal of Engineering Mechanics, Vol. 107, No. 6,
1981, pp. 1127-1238.
35 "PROBAN-The Probabilistic Analysis Program Manuals," Det
Norske Veritas Research, Rept. Nos. 89-2022, 89-2024, 89-2025, 89-2026,
89-2027, Hovik, Norway, 1989.
Purchased from American Institute of Aeronautics and Astronautics

REFERENCES 241

36 "COMPASS," Version 1.4 Manual, Martec, Ltd., Halifax, NS, Can-


ada, 1995.
37 Liu, P. L., et al., "CALREL User Manual," Structural Engineering
Mechanics and Materials Rept. Univ. of California, UCBISEMM-89/18,
Berkeley, CA, 1989.
38 Millwater, H., Wu, Y. T., Torng, T., Thaker, B., and Leung, C. P.,
"Recent Developments of the NESSUS Probabilistic Structural Analysis
Computer Program," Proceedings of the 33rd AIAA/ASME/ASCE/AHS/
A S C Structures, Structural Dynamics and Materials Conference, AIAA,
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Washington, DC, 1992, pp. 614-624.


39 WU,Y. T., "Efficient Methods for Mechanical and Structural Reliability
and Design," Ph.D. Dissertation, Dept. of Mechanical Engineering, Univ.
of Arizona, Tucson, AZ, April 1984.
40 WU,Y. T., and Wirching, P. H., "New Algorithm for Structural Reliabil-
ity Estimates," Journal of Engineering Mechanics, Vol. 115, No. 2, 1989,
pp. 2763-2781.
41 Maymon, G., "Probability of Failure of Structures Without a Closed-
Form Failure Function," Computers and Structures, Vol. 49, No. 2, 1993,
pp. 301-313.
42 Montgomery, D. C., Introduction to Linear Regression Analysis, Wiley,
New York, 1982.
43 Vanmarcke, E. H., "Stochastic Finite Element Analysis," Shinozuka
M. (Ed.) Proceedings of the Symposium on Probabilistic Methods in Engi-
neering, edited by M. Shinozuka, ASCE, New York, 1981, pp. 278-294.
44 Der Kiureghian, A., and Ke, J. B., "The Stochastic Finite Element
Method in Structural Reliability," Journal of Probabilistic Engineering Me-
chanics, Vol. 3, No. 2, 1988, pp. 83-91.
45 "ANSYS Engineering Analysis Systems User's Manuals," Swanson
Analysis Systems, Inc., Houston, PA, 1993.
46 "Design Optimization," ANSYS 4.3 Tutorials, Vol. 1, DN T0006, Swan-
son Analysis Systems, Inc., Houston, PA, 1988.
47 Nataf, A., "Determination des distributions dont les merges sont
donnees," Comptes Rendus de L' Academie des Sciences, Vol. 225, 1962,
pp. 42-43.
4X Der Kiureghian, A., and Liu, P. L., "Structural Reliability Under In-
complete Probability Information," Journal of Engineering Mechanics,
Vol. 112, No. 1, pp. 85-104.
4y Maymon, G., "Direct Computation of the Design Point of a Stochastic
Structure Using a Finite Element Code," Structural Safety Vol. 14,
pp. 185-202.
Madsen, H. O., and Tvedt, L., "Methods for Time-Dependent Reliabil-
ity and Sensitivity Analysis," Journal of Engineering Mechanics, Vol. 116,
No. 10, 1990, pp. 2118-2135.
51 Vanmarcke, E. V., "On the Distribution of the First Passage Time for
Normal Stationary Random Process," Journal of Applied Mechanics, Vol.
42, No. 3, 1975, pp. 215-222.
52 Wen, Y. K., and Chen, H. C., "On Fast Integration for Time Variant
Structural Reliability," Prob. Eng. Mech., Vol. 2, No. 3, 1987, pp. 156-162.
Purchased from American Institute of Aeronautics and Astronautics

242 G. MAYMON

53 Broek, D., Elementary Engineering Fracture Mechanics, 4th ed., Mar-


tinus, Nijhoff, The Hague, Netherlands, 1985.
54 Broek, D., The Practical Use of Fracture Mechanics, Kluwer Academic,
Norwell, MA, 1989.
55 Barsom. J. M.. and Rolfe. S. T.. Fracture and Fatime Control in Struc-
renti ice-H&
tures: ~ ~ ~ l i c a t i o n s~'roafc t u r e ~ e c h a n i c s , Englewood Cliffs,
NJ, 1987.
56 Anderson, T. L., Fracture Mechanics: Fundamentals and Applications,
CRC Press, Boca Raton, FL, 1991.
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

57 Parton, V. Z., and Morozov, E. M., Mechanics of Elastic-Plastic Frac-


ture, Hemisphere, New York, 1988.
58 Paris, P. C., and Erdogan, F., "A Critical Analysis of Crack Propagation
Laws," Journal of Basic Engineering, Vol. 85, No. 4, 1963, pp. 528-534.
59 Bogdanoff, J., "A New Cumulative Damage Model," Journal of Ap-
plied Mechanics, Pts. 1-3, Vol. 46, 1979, pp. 245-257, 733-739.
60 Bogdanoff, J., and Kozin, A., "A New Cumulative Damage Model,"
Journal of Applied Mechanics, Pt. 4, Vol. 47, 1980, pp. 40-44.
61 Sobczyk, K., and Spencer, B. F., Random Fatigue, From Data to Theory,
Academic, New York, 1992.
62 Noronha, P. J., et al., "Fastener Hole Quality, I and 11, U.S. Air Force
Flight Dynamics Lab. Tech. Rep. AFFDL TR-78-206, Wright-Patterson
AFB, OH, 1979.
63 Virkler, D. A., et al., "The Statistical Modelling Nature of Fatigue
Crack Propagation," Journal of Engineering Materials and Technology,
Vol. 10, NO. 4, 1979, pp. 148-153.
Ghonem, H., and Dore, S., "Experimental Study of the Constant Proba-
bility Crack Growth Under Constant Amplitude Loading," Engineering
Fracture Mechanics, Vol. 27, No. 1, 1987, pp. 1-25.
65 Murukami, Y., et al. (eds.), Stress Intensity Factors Handbook, 3rd ed.,
Vols. 1 and 2, Pergamon, Oxford, England, UK, 1990.
66 Cherepanov, G. P., Mechanics of Brittle Fracture, McGraw-Hill, New
York, 1979.
67 Lin, Y. K., and Yang, J. N., "A Stochastic Theory of Fatigue Crack
Propagation," A I A A Journal, Vol. 23, No. 1, 1984, pp. 117-124.
Lin, Y. K., Wu, W. F., and Yang, J. N., "Stochastic Modeling of Fatigue
Crack Propagation," Proceedings of the IUTAM Symposium on Probabilis-
tic Methods in Mechanics of Solids and Structures, Springer-Verlag, Berlin,
1984, pp. 103-110.
69 Sobcyzk, K., "Stochastic Modeling of Fatigue Crack Growth," Proceed-
ings of the IUTAM Symposium on Probabilistic Methods in Mechanics of
Solids and Structures, Springer-Verlag, Berlin, 1984, pp. 111-119.
O' Spencer, B. F., and Yang, J., "Markov Model for Fatigue Crack
Growth," Journal of Engineering Mechanics, Vol. 114, No. 12, 1988,
pp. 2134-2157.
Yang, J. N., Salivar, G. C., and Annis, C. G., "Statistical Modeling of
Fatigue Crack Growth in Nickel Based Superalloy," Engineering Fracture
Mechanics, Vol. 18, No. 2, 1983, pp. 257-270.
Purchased from American Institute of Aeronautics and Astronautics

REFERENCES 243

72 Sobczyk, K., "Modeling of a Random Fatigue Crack Growth," Engi-


neering Fracture Mechanics, Vol. 24, No. 4, 1986, pp. 609-623.
73 Zhu, W. Q., Lin, Y. K., and Lei, Y., "On Fatigue Crack Growth Under
Random Loading," Engineering Fracture Mechanics, Vol. 43, No. 1,
1992, pp. 1-12.
74 Libiard, A. B., "Probabilistic Fracture Mechanics," Fracture Mechanics,
Current State, Future Prospects, Pergamon, Oxford, England, UK;1979.
75 Varanasi, S. R., and Wittaker, I. C., "Structural Reliability Prediction
Method Considering Crack Growth and Residual Strength," Fatigue Crack
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

Growth Under Spectrum Loads, STP 595, ASTM, Philadelphia, 1976,


pp. 292-305.
76 Ortiz, K., "Stochastic Modeling of Fatigue Crack Growth," Ph.D. Dis-
sertation, Dept. of Civil Engineering, Stanford Univ., Stanford CA, 1984.
77 Ortiz, K., and Kiremidjian, A. S., "Time Series Analysis of Fatigue
Crack Growth Rate Data," Engineering Fracture Mechanics, Vol. 24, No.
5, 1986, pp. 317-334.
78 Ortiz, K., and Kiremidjian, A. S., "Stochastic Modeling of Fatigue
Crack Growth," Engineering Fracture Mechanics, Vol. 29, No. 3, 1986,
pp. 317-334.
79 Lawrence, M., Liu, W. K., Besterfield, G., and Belytschko, T., "Fatigue
Crack Growth Reliability," Journal of Engineering Mechanics, Vol. 116,
NO. 3,1990, pp. 698-708.
80 "Fatigue Crack Growth Computer Program NASAIFLAGRO," Ver-
sion 2, NASA Publ. JSC-22267A, Houston, TX, May 1994.
Yang, J. N., and Donath, R. C., "Statistical Crack Propagation in
Fastener Holes Under Spectrum Loading," Journal of Aircraft, Vol. 20,
NO. 12,1983, pp. 1028-1032.
82 Tanaka, H., and Tsurui, A., "Reliability Degradation of Structural
Components in the Process of Fatigue Crack Propagation Under Stationary
Random Loading," Engineering Fracture Mechanics, Vol. 27, No. 5,
1987, pp. 501-516.
83 Tsurui, A., Nienstedt, J., Schueller, G. I., and Tanaka, H., "Time Vari-
ant Structural Reliability Analysis Using Diffusive Crack Growth Models,"
Engineering Fracture Mechanics, Vol. 34, No. 1, 1989, pp. 153-167.
84 Larsson, H., and Bernard, J., "Fracture of Longitudinally Cracked
Ductile Tubes," International Journal of Pressure Vessel Piping, Vol. 6,
1978, pp. 223-243.
85 Feddersen, C. E., "Evaluation and Prediction of the Residual Strength
of Center Cracked Tension Panels," STP 486, ASTM, Philadelphia,
1970, pp. 50-78.
Acoustic Fatigue Sub-series Vol. 1, Data Sheets 66013, 66016, 66017,
66018, ESDU International, London, 1976, 1978.
Purchased from American Institute of Aeronautics and Astronautics
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

This page intentionally left blank


Purchased from American Institute of Aeronautics and Astronautics

ANSYS, 77, 81, 148, 151, 153, 157, 163, First-order reliability method (FORM),
165-167, 171-172, 174, 177-178 115-117, 123, 133, 140, 142, 154,
214215
Base excitation, 4-6, 2 6 2 9
FLAGRO, 200
Beam, simply supported, 32,68-71.98, 188
Fracture mechanics, 195
Boundary-layer excitation model, 43-51
Burdekin-Stone criterion, 210-211 Gaussian excitations, 84-86, 103
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

General force input, 6 1 1


CALREL, 133-134, 193,215
Geometric nonlinearity, 82
Cantilever beam, 31, 140, 147, 165
Chen-Lind approach, 120 Harmonic excitations, external, 1-4
Choleski decomposition, 163 Hasofer-Lind method, 118, 123
Classical oscillator, 1 Hoenbichler-Rackwitz approach, 120
COMPASS, 133
Composite structures, 148 Importance factors, 132- 133
Cornell reliability index, 117-1 18 Iterative procedure, 104, 131-132
Crack growth, 195-197,210,213,216
Lagrange multiplier method, 124-130, 142,
Crack growth models, 195-196
147, 151, 154-155
Crack growth rate, 197-199,201-203,205,216
Linear structures,62-68
Crack length, 198,202-209,212-214,217
Linear systems, 5 3 4 2
Crack size, 196-198, 209
Load cycles, 196197
Crocker model, 44
Cumulative distribution function (CDF), Maestrello's experiments, 44-45
35-36, 112-113, 118-122, 128, 158, 170, Markov-chain crack growth model, 202
182-183 Microstructures, 196
crack length, 206-208 Modal damping coefficients, 60,88-89.95
Modified joint probability density function
Design point, 157, 159, 162, 165, 173, 178
(MJPDF), 157-163
Deterministic continuous system
numerical examples, 163-178
base excitation, 26-29
nonnormal dependent variables, 169-173
differential equations, 22-26
normally distributed, correlated variables,
stress response, 29-34
167-168
Deterministic loads, 25
normally distributed, random field,
Differential equation models, 197, 199
correlated variables, 174-178
Dufting oscillator, 74
normally distributed, uncorrelated
Duhamel integral, 7, 9 variables, 164-166
Dynamic load factor (DLF), 3, 8, 29 threshold value problem, 173-174
Dynamic response, 82 Monte Carlo simulation, 108, 133, 142-143,
Elastic bar, 108-109 154155
Elastic system, 98-103 Multiple-degree-of-freedom(MDOF) systems,
Excitation functions, 41-43 21, 34, 6365.67, 79
differential equations 13-15
Failure criterion, 209-21 1,216217 equilibrium equations, 13
Failure surface, 110-1 17, 124, 130, 135, 142, generalized masses, dampings, rigidities,
I48 and forces, 15-18
Fatigue, 195 nonlinear random response, 98-100
Finite element method, 21,25, 31, 33, 72-73, normal modes, 13-16
77,79,82,98, 104, 123, 133-135, 148, random excitations, 56-62, 184
157, 159-161, 163-164, 166, 174-175, spectral moments, 1 8 4 185
177-178,215 uncoupled differential equations, 18-19
Purchased from American Institute of Aeronautics and Astronautics

246 INDEX

Nataf model, 160, 162 Safety factor, 107-109


NESSUS, 134 Second-order reliability method (SORM),
Nondeterministic structures, 107-108, 179 116-117, 133, 140, 142, 154-155.214-215
Nonlinear coefficients, 79-82 Sensitivity, 132-133
Nonlinear response Single-degree-of-freedom (SDOF) systems,
elastic systems, 98-103 34.42, 62,67
SDOF systems, 87-92 base excitation, 4-6
two-DOF systems, 92-98 external harmonic excitation, 1-4, 56
Nonlinear structures, 73 general force, 6-1 1
Normal modes, 13-16.21, 24, 30, 33,94,98, nonlinear behavior, 74-79, 87-92
Downloaded by RMIT UNIV BUNDOORA on July 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/4.866494

102 random excitations, 53-56


Numerical programs, 132-134 random Gaussian force, 87-92
stresses, 11
Panel vibration, 43 transfer function, 55
Periodic loads, 195 S-N curves, 195
Power spectral density (PSD) functions, Sound pressure level (SPL), 41,47
4142,45,47, 51, 54-56.58, 61,68, Stationary Gaussian process, 179-183
179-180, 184, 186, 189 Statistical linearization method
Pressure excitations, 47, 49 MDOF systems, 83-87
Pressure fluctuations, 43 Stochastic crack growth models, 196202
Probabilistic analysis, 107-108, 110, differential equation models, 197, 199
132-135, 157-158 Stresses, 11
Probability density function (PDF), 35-36, Stress intensity factor, 215
112-113, 115, 118-121, 123, 138, 140, 142, Stress modes, 30-34, 70-71, 104
170, 182-183 Stress response calculation, 104-106
crack length, 207-209, 21 3-2 14.2 17 Structural analysis, 107
Probability of failure, 123-124, 151, 165, Structural design, 107, 210
178, 182, 193, 197.211-215.217
Probability of threshold crossing, 185-193 Taylor series expansion method, 123,
PROBAN, 133, 142-143, 146-147, 154-155, 135-137, 157, 188, 190-191
165-167, 175, 178 evaluation point selection, 137-140, 143,
145, 149-153
Rackwitz-Fiessler method, 118-121, 123, examples, 140-155
131 Tensile stress, 109, 153
Random excitations, 35, 38-39,84 Threshold value problem, 173-174
linear structures, 6 2 4 8 Two-degree-of-freedom(two-DOF) systems,
practical characterization, 3 9 4 1 5 8 4 2 , 6 7 , 184
SDOF systems, 76 random Gaussian force, 92-98
stationary, 54-56, 6 2 4 8 Turbulent boundary layer, 43
Random functions, 53 Turbulent flow, 38,43
basic concepts, 35-39 autocorrelation functions, 5 1
Random Gaussian force, 87-98 Uncertainty approach, 107
Random process (RP) method, 197,2W202,
205, 217 Vibrating structures, structural vibration, 1, 8,
Random structures, 219 194
Random variable (RV) methods, 197, Viscous damping, 1,92
199-200,205 von Mises formulation, 140
Random vibrations, 1, 8, 219
Rectangular plates, 32-34,49 Weibull distribution, 129, 131, 169
White noise, 4 1 4 2
Reliability, structural, 107, 127
Wideband excitation, 40
Reliability index, 117-122, 140, 143, 151,
Wiener-Khintchine formulas, 37.43, 49, 54,
162, 165, 173, 177
58,66
Reliability methods, 115-1 17
Wind tunnel measurements, 43
Resonance frequencies, 19.21, 24, 56, 89
Rosenblatt transformation, 120 Yield stress, 129, 131

Potrebbero piacerti anche