proppe03_equivalent linearization and mc
TRANSCRIPT
-
8/10/2019 Proppe03_Equivalent Linearization and MC
1/15
Equivalent linearization and Monte Carlo simulationin stochastic dynamics
C. Proppe*, H.J. Pradlwarter, G.I. Schueller
Institute of Engineering Mechanics, University of Innsbruck, Technikerstr. 13, A-6020 Innsbruck, Austria
Abstract
Equivalent linearization (EQL) and Monte Carlo Simulation (MCS) are the most important techniques in analyzing large nonlinear
structural systems under random excitation. This paper reviews the development and the state-of-the-art of EQL and MCS in stochastic
structural dynamics.q 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Complex modal analysis; Controlled Monte Carlo simulation; Importance sampling; Linearization criteria; Local linearization; Neumann
expansion
1. General remarks
An extensive benchmark study [113] has revealed that
most of the computational methods developed in stochastic
structural dynamics are quite limited with respect to thedimension of the problem and consequently cannot be
applied to engineering problems of larger size. It appeared
that only two approaches, namely equivalent linearization
(EQL) and Monte Carlo simulation (MCS), are suited to
solve these problems. Other methods, such as numerical
solutions based on the FokkerPlanck equation (FPE) by
path integration[94,136], cell-mapping[60,61]and finite-
elements[14,128], increase in complexity at least exponen-
tially with the state-space dimensionn. In fact, all numerical
solutions of the FPE available in the literature belong to
low-dimensional problems with n # 4[140].
Both MCS and EQL are conceptional simple and easy toapply. There underlying concepts can be viewed as the poles
between which most of the computational methods devel-
oped so far in stochastic dynamics can be settled: EQL, on
the one hand, uses the simplest global approximation of the
distribution of the response, while MCS is the most local
approach without any a priori assumption. Whereas only
biased estimates for the first two moments of the response
are obtainable by the standard EQL procedure, MCS yields
unbiased estimates for the probability density function of
the nonlinear response. Moreover, the information that EQL
provides is limited to the first two moments of the
distribution of the response, while MCS techniques are
able to give additional information on the tails of the
distribution and subsequently credible reliability infor-
mation on the analyzed dynamical systems.
2. Equivalent linearization
2.1. Introduction
An EQL technique for random external excitation aims
to replace the n-dimensional nonlinear system
dXt fX; tdtdFt; 1
where fX; t is the nonlinear system function and Ft theexcitation process, by a linear system
dXt AXadtdFt; 2such that a certain approximation error is minimized. The
n n-dimensional matrixA and then-dimensional vectora
contain the linearization coefficients. A frequently used
error criterion is the difference
efx; t2Ax 2 a; 3between the system functions which is minimized in mean
square sense. After some mathematical manipulations, this
0266-8920/03/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 2 6 6 - 8 9 2 0 (0 2 )0 0 0 3 7 - 1
Probabilistic Engineering Mechanics 18 (2003) 115
www.elsevier.com/locate/probengmech
* Corresponding author. Tel.: 43-512-507-6843; fax: 43-512-546-865.
E-mail address: [email protected] (C. Proppe).
http://www.elsevier.com/locate/probengmechhttp://www.elsevier.com/locate/probengmech -
8/10/2019 Proppe03_Equivalent Linearization and MC
2/15
leads to the equations
aEfX2AEX 4and
EfiXX2 EfiXEX
EX2 EXX2 EXT
ATi ; i1; ; n
5
for the linearization coefficients, where Aiis the ith row of
matrixA.
EQL has been originated independently by Kazakov,
Booton and Caughey. Kazakov and Booton, who were
concerned with problems of control theory and automatiza-
tion, considered the linearization of nonlinear memoryless
transformations of random variables (RVs). While the work
of Booton seemed to have only a small impact, the
linearization technique of Kazakov was applied and
extended by several authors in the Soviet-Union. Both
authors applied the method also to nonlinear systems under
Gaussian white noise excitation. In this case, an additionaldifficulty occurred: in order to calculate the linearization
coefficients, the probability density function of the nonlinear
system is needed. In general, thisprobability density function
is unknown. Therefore, approximations of the true
linearization coefficients were calculated by replacing the
probability density function of the nonlinear system with the
probability density function of the linearized system. This
has been done under the assumption that the expectations
needed for the calculation of the linearization coefficients
will not change very much when they are evaluated with
respect to the probability density function of the linearized
system. This substitution necessitates successive approxi-mations in order to calculate the approximate linearization
coefficients.
Caughey developed EQL when he tried to generalize the
EQL technique invented by Krylov and Bogoliubov. He
investigated the Van-der-Pol oscillator[25], vibrations of a
nonlinear string [24] and a system with hysteresis [26].
Furthermore, within the development of EQL, it was
important to recognize that for a Gaussian probability
density function, the following relation holds:
EfiXX EXXTE7XfiX: 6Using this formula, the linearization coefficients are obtained
from the expectation of the gradient of the nonlinear
function. This result has been published in Refs. [6,76].
Further simplifications of practical importance are possible
using the linearity property of the expectation operator, e.g.
for chain-like systems[48].
The above outlined procedure may be easily extended to
nonstationary problems, where the linearization coefficients
vary with time. EQL is then carried out in successive discrete
points on the time-axis. The expectations must now be
obtained for a linearsystemwith time-varying coefficients. In
case of Gaussian excitation, the expectations may be
obtained from the solution of differential equations for the
mean value and the covariance matrix[111, p. 113].In the
iterative procedure for the determination of the linearization
coefficients, the solution of Eqs. (4) and (5) alternates then
with an integration of the differential equations for each time
interval [100]. The length of the time intervals can be
controlled by comparing results obtained with different step
sizes.For non-Gaussian excitation, the evaluation of the
expectations may be much more involved. In Refs. [44,
52,66,134], EQL has been applied to systems under
external Poisson white noise excitation. In Refs. [52,134],
the Duffing oscillator is considered, in Ref. [44] a uni-
dimensional system with cubic nonlinearity and in Ref.
[66] a SDOF system with Bouc Wen hysteresis. For
polynomial type nonlinearities, the response moments of
the linearized system have to be computed. This has
been done in Ref. [134] for SDOF systems and in Ref.
[52] for MDOF systems. In case of nonpolynomial type
nonlinearities, the probability density function of the
linear response is approximated by a GramCharlier
series A in Ref. [66]. Looking closer at the results of
these papers, one observes that in most of the cases the
stationary standard deviation of the response differs only
slightly from that obtained by a corresponding Gaussian
excitation. Thus, considering a model with Gaussian
excitation might be an interesting alternative in case that
the random impulses are not relatively rare events.
2.2. Theoretical developments
2.2.1. Linearization criteria
Different linearization techniques may be definedaccording to different definitions of the approximation
error. An approximation error envisages often a specific
application and bears some subjectivity. A large part of
the literature on EQL is devoted to the development and
application of new linearization criteria. However, in
practice only the mean square criterion has found
widespread application.
Kazakov [75] introduced the equality of the first and
second order moments of the linearized and the nonlinear
system function as an error criterion. He also proposed to
take the arithmetic mean of the linearization coefficients
that were calculated according to this error criterion and
the mean square criterion. Criteria based on the potential
and dissipative energy were proposed in Ref. [41].
These criteria are compared for the dimensionless
Duffing oscillator under Gaussian white noise excitation
dX1t X2tdt; 7
dX2t 2cX2tdt2X1t1 eX21t ffiffiffi
2cp
dWt:This equation is linearized by
dX1t X2tdt;dX2t 2cX2tdt2X1t1k ffiffiffi2c
p dWt;
8
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 1152
-
8/10/2019 Proppe03_Equivalent Linearization and MC
3/15
where the linearization coefficient k has the following
form
minimization of difference between system functions:
k1eEX41t=EX21t minimization of difference between potential energies:
k2eEX61t=2EX
41t
equality of mean square system functions: k3effiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
EX61t=EX21tq
equality of mean square potential energies: k4eEX41t=2EX21t
Especially, it is seen that k12k4: Evaluating theexpectations with respect to the probability density function
of the linearized system, the stationary standard deviation of
the displacement is obtained in the form sX1 d=e1=4: Theasymptotic behavior for large e of the approximated and
exact stationary standard deviation corresponds, as one can
show thatsX1!
0:8222
=e
1=4
[111]. The asymptotic error issummarized in the last column ofTable 1. In this case, EQL
yields good approximations also for large nonlinearities [34].
It is noted that except for the equality of the mean square
potential energies, the approximations of the standard
deviation are unconservative.
Polidori and Beck [99] chose the linear system whose
stationary probability density function minimizes the FPE
operator of the nonlinear system with respect to the L2 norm
and a weighted L2 norm, respectively. For the Duffing
oscillator and the Duffing oscillator with cubic damping
they improved the results obtained with the mean square
criterion.
In order to calculate power spectral density approxi-mations, Apetaur[24]developed linearization techniques
that are based on statistical characteristics of second order,
especially on the correlation function. Iyengar[67]tried to
approximate the nonlinear system by a linear system of
higher order. The additional linearization coefficients are
determined from time-derivatives of the nonlinearity. This
leads to additional differential equations with parametric
excitation. In this way, Iyengar intended to approximate
higher resonances.
Bernard and Taazount [16] investigate systems, where
the nonlinearity is due to springs with a gap. They show that
the usual linearization technique will yield goodapproximations of the power spectral density, if the gap is
very small or very large compared to the standard deviation
of the system response. If this is not true, then the response
spectrum contains a broad band. In this case Bernard and
Taazount propose to approximate the power spectral density
by the power spectral density of an ARMA process. Then, at
least an approximation of the power spectral density must be
given.
With the aim to apply EQL to reliability calculations,Casciati et al.[23]compare the stationary level crossing rate
at some critical level. They obtained a local linearization
technique that has been successfully applied to the Duffing
oscillator and to hysteretic systems, respectively.
Recently, several authors proposed to calculate the mean
square error with respect to the probability density function
of the linearized system [40,124]. In this case, the derivatives
with respect to the parameters in the Gaussian probability
densityfunction has to be taken into consideration. However,
the results that were obtained with this linearization
technique were worse than those obtained from the usual
mean square criterion.The reasonfor this will be explained in
Section 2.2.2. Unfortunately, by calling the proposed method
as corrected linearization[124]and the usual method as
erroneous [40] the authors gave the impression that the
usual method is false. It is once more mentioned that
the minimization procedure is carried out with respect to the
probability density function of the nonlinear system and that
at a last step, this probability density function is approxi-
mated by the probability density function of the linearized
system. The approximation error is thus not minimized with
respect to the probability density function of the linearized
system. For a further discussion of this topic, the reader is
referred toRef. [36].
2.2.2. Approximation error
Only a few theoretical papers are devoted to the
estimation of the approximation error. Most authors study
specific systems and obtain the approximation error by
comparison with exact solutions, simulation results or with
cumulant or quasi-moment closure methods. For external
Gaussian white noise excitation, comparison with results
from closure techniques is in so far justified as Gaussian
closure and EQL may be viewed as equivalent approaches.
Thus, higher order closure schemes can be considered as
higher order corrections of the results obtained by EQL.
However, for MDOF systems, higher order closure schemes
are very time demanding such that in this case, only
simulation techniques may be suited for error estimation.
The results from comparative studies[111], show that EQL
may give accurate approximations of the standard deviation
of the response even if the nonlinearities are not small. For
MDOF systems the error may be higher [92]. However, in a
recent benchmark study [113], itis concluded thatEQLis
at least for the predictionof the variance sufficiently robust
accurate and also applicable with an acceptably low effort.
In general, the standard deviation is underestimated.
The following remarkable result has been pointed out by
Caughey [27], Crandall [35] and Kozin [81]: if
Table 1
Asymptotic behavior of the standard deviation of displacement for different
linearization criteria
Linearization criterion d % Error
Minimization of difference between system functions 0.7598 7.6
Minimization of difference between potential energies 0.7953 3.3
Equality of mean square system functions 0.7128 13.3
Equality of mean square potential energies 0.9036 9.9
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 115 3
-
8/10/2019 Proppe03_Equivalent Linearization and MC
4/15
the linearization coefficients were obtained from expec-
tations with respect to the probability density function of the
response of the nonlinear system, then the first and second
order moments of the linearized and the nonlinear system
would coincide. The error is thus introduced by the
replacement of the probability density function of
the nonlinear system by the probability density functionof the linearized system.
Several authors tried to improve the probability density
function that is used to obtain the linearization coefficients.
Beaman and Hedrick [13] approximate the probability
density function by a GramCharlier series A of fourth
order. Chang [29] assumes that the probability density
function is a sum of weighted Gaussian probability density
functions. Hurtado and Barbat[62]represent the probability
density function by a Gaussian probability density function
and a Dirac delta function. All these methods are character-
ized by a higher computational effort, especially for MDOF
systems. However, the effort that is neededfor thecalculation
of these approximations may not always be justified by the
accuracy of the results.
Kolovskii [79] estimates the standard deviation of the
process Xt2X0t; where X0t is the state vector of thelinearized system. These estimates contain the standard
deviation of the nonlinearity for which a Lipschitz condition
is assumed to hold.
Skrzypczyk[122]utilizes the theory of integral equations
over locally compact abelian groups in order to estimate a
measure of the difference Xt2X0tby a measure of thedifference between the nonlinear and the linear system
function. In this way he justifies the use of some
linearization techniques.Bernard[15]and Bernard and Wu [17], making use of
the large deviation principle, give also a justification of some
linearization techniques. They consider the norm of the
difference between the probability measures of the non-
linear and the linearized system and estimate it by the
relative entropy, which is then minimized.
2.3. Extension of EQL
2.3.1. EQL for parametric excitation
For parametric excitation by Gaussian white noise,
different linearization techniques may be obtained byconsidering different models for the linearized system. The
following linear models were proposed in the literature:
linearized parametric excitation[18,31]
external excitation instead of parametric excitation [43,
80,123,141]
equivalent linear restoring forces instead of parametric
excitation [129]
In this context, it has been shown that models that retain
the parametric excitation may yield even less accurate first
and second response moments than other models, although
only linear parametric excitation is present in the original
nonlinear problem. Bruckner and Lin[18]pointed out that
the linearized system may not represent the stability
properties of the nonlinear system. Especially for systems
with bifurcation points, equivalent nonlinearization may
lead to a better representation of the stability properties[80].
The linearization procedures itself can be based ondifferent arguments:
linearization of the equations of motion[31]
linearization of the Itoequation[18,43,123]
linearization of the Ito formula[43].
Several proposed linearization techniques have the
property that the truly linearized system and the nonlinear
system have the same first and second order moments, see
Ref. [107]for details.
2.3.2. Application to hysteretic systems
For systems with hysteresis, the nonlinear systemfunctions depend on the history of the system response.
Caughey [26] proposed the application of the averaging
principle [82] in order to apply EQL to systems with
hysteresis. The approach is based on the assumption that the
response has a narrow frequency band. After introduction of
an amplitude and a phase process, the expectations can be
obtained after averaging by integration with respect to the
amplitude distribution. This method has been further
developed and investigated inRefs. [50,78,131]. However,
as has been shown inRefs. [64,109,137], the response may
also contain a broad frequency band. In this case, the
standard deviation is considerably underestimated by EQL.Another approach, due to Wen [137], considers the
introduction of an additional nonlinear differential equation
in order to account for the hysteretic behavior. The system
function is given by
fx;_x c_x akx2 12 akzx; 9where the hysteretic part zxis described by the first orderdifferential equation
_z bl_xllzln21z g_xlzln K_x; 10with constants b, g, K and n. These constants may be
adapted such that a wide variety of hysteresis loops may be
described. The equations of motion together with the
differential equations for the hysteretic part form a nonlinear
system of differential equations, which is then linearized.
Extensions of the so-called Bouc Wen model can be found
in Refs. [10,38,47,98,139]. An overview of the obtained
results is contained in the review article [138]. Despite its
broad acceptance, the BoucWen model is not in agreement
with classical plasticity theory. Extensions of the model to
overcome these contradictions can be found inRef. [20].
In order to treat any kind of nonlinearity, including
hysteretic material laws, a simulation procedure is proposed
in Ref. [100]. The distribution of the state vector of
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 1154
-
8/10/2019 Proppe03_Equivalent Linearization and MC
5/15
the linearized system is simulated and the linearization
coefficients are obtained from a least squares problem. For
Gaussian excitation, the approach requires the generation of
a Gaussian distributed RV only. In case of impulsive
excitation (e.g. by Poisson white noise), the state vector can
be simulated easily and without much numerical effort.
2.3.3. Application to continua
In order to analyze continua, a discretization of the
problem, e.g. by the FE method, is necessary. EQL can be
applied to the differential equations of the continuum or to
the differential equations of the discretized problem.
Iwan and Krousgrill [63] consider the first possibility,
which has the advantage that the material laws and the yield
surface can be specified for the continuum[22, S. 287]. In
Refs. [93,121], continua under nonstationary Gaussian
excitation with visco-elastic properties described by the
BoucWen model are investigated. The authors apply EQL,
discretize the structure and neglect higher modes of thelinearized system. Casciati and Faravelli [21] consider
elasto-plastic material properties with yield surfaces, hard-
ening rules and hysteresis. They also employ a BoucWen
model and linearize the equations prior to discretization.
Iwan and Whirley [65] introduce a decomposition for the
field of linearization parameters for nonlinear continua
under nonstationary random excitation. The minimization
conditions are obtained from a variational problem.
The linearization of the discretized equation of motion
has been proposed for beams[19,118]and plates [1]. The
advantage of this approach is that the discretized system is
built only once, which is extremely useful when coupling
EQL with commercial FE programs.
2.3.4. Local linearization
The need for increasing the accuracy of the results
obtained by EQL led many researchers to extensions of
EQL. For external excitation, one can show that EQL and
Gaussian closure of the moment equations are equivalent
approaches. Most closure schemes can therefore be viewed
as higher order extensions of EQL. These extensions suffer,
however, from the fact that they are difficult to apply and
that the computational efforts increase tremendously for
MDOF systems. One of the reasons for these facts lies in the
global approximation concept applied in EQL and its
extensions.
A local linearization approach has been recently
suggested in Ref. [104], where the nonlinearity is locally
linearized. As a result, the probability density function of
the nonlinear system is very accurately approximated by
suitable Gaussian probability density functions. In this way,
a local linearization approach may be suited, e.g. for
reliability problems, where the tails of the probability
density function have to be approximated accurately. In
contrast to closure schemes, the introduction of Gaussian
probability density functions in the phase space provides
a link between linearization methods and sampling
techniques such as MCS.
2.4. Efficient implementation of EQL for FE models
2.4.1. Local coordinate systems
Dealing with nonlinear structures discretized by FE, thenonlinearities are represented by nonlinear elements.
Linearizing these nonlinear elements employing EQL, the
linearized elements must be determined separately for each
nonlinear element as it is also done in the standard case
where all elements are built independently and then
assembled to the global matrices.
Let ug denote the generalized degrees of freedom
(DOF) of the structure and let eug , ug be the DOF
comprising only components associated with a nonlinear
element. The EQL procedure results then in structural
matrices (stiffness, damping, auxiliary variables) associ-
ated with the DOF of eug: Each linearized element must
be free of stress under rigid body motion, and is
therefore singular. For example, the stiffness matrix of a
linearized spring has a rank of one, although the matrix
is generally of size six by six within a three-dimensional
model. These specific properties of elements suggest that
the linearization should not be performed in the global
coordinate system involving all the DOF of eug but in a
local coordinate system eul involving only the minimal
number of DOF [42]. In this way, the efficiency is
enhanced and ill-conditioned problems are avoided. Also,
the nonlinear restoring force of the elements is defined
best in local minimal coordinates, i.e. for the sake of
clarity and uniqueness. Hence, the restoring force can besafely assumed to be defined in local coordinates efleul:A linear transformation
eul e Teug; 11
with constant transformation matrix eT may be intro-
duced to reduce the linearization to a smaller set of
coordinates. For the statistical linearization the first two
moments of the stochastic response will be needed in
local coordinates. The mean eml and covariance matrixeSl are readily determined from the mean
emg and
covariance matrix eSg of the element response in global
coordinates using the transformations eml
e Temg
andeSl e TeSgeTT: The linearization coefficients are thencomputed in the minimal local coordinate system and
subsequently transformed to the global coordinate system
employing the general transformation rules of element
matrices in FE analysis.
2.4.2. Stochastic response of linearized system
Linearizing a structural nonlinear model in context of
general FE-models, some peculiarities must be con-
sidered by computing the linearized response to stochas-
tic excitation. The stiffness and damping elements
resulting from EQL are in general not symmetric and
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 115 5
-
8/10/2019 Proppe03_Equivalent Linearization and MC
6/15
hence ruling out the standard equation solvers for linear
systems (which assume symmetry of the structural
matrices). Modeling hysteretic behavior may require
additional auxiliary variables, where the associated
linearization coefficients are again nonsymmetric. More-
over, modeling the excitation by filtered white noise
requires augmented structural matrices.The stochastic analysis determines the mean and
covariance matrix of the stochastic response. Two basic
approaches are currently available. The first employs
complex modal analysis and the second direct step-by-
step integration schemes.
2.4.2.1. Complex modal analysis. Complex modal analysis,
as originally suggested for MDOF-systems [30], is not
well suited to treat the system of equations resulting
from the linearization of FE-models with a large number
of DOF. Since the matrices have not the standard form,
the highly optimized FE-algorithms for symmetric
classical eigensolutions are not applicable. A state
representation is generally used, involving the displace-
ment, velocities and auxiliary variables, leading to
nonsymmetric matrices with a size . 2m; where m
denotes the number of DOF. The solution of the
nonclassical eigenvalue problem is computationally
quite expensive compared with the classical form of
size m. Moreover, dealing with nonstationary problems,
the nonclassical complex eigenvalue problem has to be
solved numerous times, increasing the computational
burden such, that the analysis of larger FE-systems
becomes intractable. Most of the above mentioned
difficulties can be circumvented by the approach shownin Ref. [115]. It is based on the solution of the classical
eigenvalue system for the symmetric part of the
structural matrices and an extended form of component
mode synthesis [33,34]. However, this approach is
certainly not practical for the solution of nonzero mean
problems, particularly when comparing its efficiency with
the recently developed direct step-by-step integration
schemes.
2.4.2.2. Direct step-by-step integration. Stochastic versions
of special integration schemes have been suggested [91,
97,133]. All these proposed procedures operate explicitly
with the covariance matrix and require special integration
schemes. Since the covariance matrix is fully populated
and of a size . 2m; these approaches cannot be applied
efficiently to FE-systems with a large number m of DOF,
because of storage requirements and number of compu-
tational operations.
A solution technique which avoids the storage and
direct integration of the full covariance matrix has been
developed recently in Refs. [105,106], where it is
suggested that the covariance matrix is represented by
the Karhunen Loeve expansion with a reduced number
of Karhunen Loeve vectors. Continuous white noise is
approximated by discrete noise (shot noise, random
impulse) applied at one instant in each time step. The
KarhunenLoeve vectors can be integrated by any
available and appropriate deterministic step-by-step
integration scheme. These vectors are further updated
in each time-step to take the effect of the random
impulse loading into account. Most important for theefficiency of the suggested approach is the fact that the
update can be carried out exactly in a low-dimensional
reduced subspace. Although this approach requires only
deterministic direct integration schemes [11], these
integrations must be performed NK-times, where NKdenotes the number of Karhunen Loeve vectors. Since
NK lies in the range of 20200, depending on the
problem and the required accuracy, the computations
might be still a tedious task, particularly for very large
FE-systems. For nonlinear systems with large linear
subsystems, modal transformations can be used for a
considerable size reduction of the FE-model. Its appli-
cation in context with nonlinear structural stochastic
analysis is shown in Ref. [106] including the effects of
mode truncation.
2.5. Areas of application
As many comparative studies revealed, EQL yields in
general good approximations for the first and second
order moments of the system response. For most
practical applications, e.g. reliability analyses, the
information provided by EQL is not sufficient. Thus, it
has been often argued that EQL should be employed
during the design stage, in order to obtain a first estimateof the system properties. There seems to be no
systematic studies concerning the applicability of EQL
to design problems. Therefore, possibilities of application
for EQL during the design stage are outlined here,
namely sensitivity analyses (e.g. in order to reduce the
simulation efforts for systems containing random par-
ameters) and the design of controlled structures.
2.5.1. Sensitivity analysis
Suppose that in Eq. (1), the state vector and the system
functions dependent on a design parameter b:
dXt; b fXt; b; bdt FbdWt: 12Taking the first derivative of the state vector with respect to
the design parameter leads to
dZt; b fb
X; bdt fxi
ZidtdF
dbdWt; 13
here Zdx=db: The differential equation (12) acts as afilter for Eq. (13). In order to obtain approximations for
the mean and variance of the design sensitivity with
respect to the parameter b, the application of EQL has
been proposed in Ref. [108]. It is observed that the
original system (12) can be linearized independently of
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 1156
-
8/10/2019 Proppe03_Equivalent Linearization and MC
7/15
Eq. (13). The linearization coefficients for Eq. (12) can
be used for the linearization of Eq. (13).
The linearization of Eq. (13) leads to
dZt; b A1bXt; bdtA2bZt; bdt a3bdt
F
bdW
t;
14
with
A1ijb E
2fi
xjb
2fi
xjxkZk
" #
A2ijb E fi
xj
" #Aij; 15
A3ib E fi
b
E fi
xjZj
" #2A1ijEXj2AijEZj:
The solution scheme consists of two fundamental steps:
1. Solve for the linearization coefficients A, a of system
(12).
2. Minimize A1ijXAijZa3i 2 fi=b2 fi=xkZk inmean square sense and calculate the expectations EZ;EZZT:
A further reduction of the computational effort would be
possible, if the linearized system, together with the system
of differential equations for the derivative of the state vector
with respect to the design variables, would lead to a closed
set of equations for higher order moments. However, this is
in general not the case, even if the nonlinearities in theoriginal system are of polynomial type. Thus, the compu-
tational effort can be reduced from the linearization of a
2n 2n-dimensional system to the linearization of two
n n-dimensional systems.
2.5.2. Stochastic structural control
Control methods for linear systems under external
Gaussian excitation have been investigated by many
researchers. In most of the studies, linear feedback control
and mean square cost functionals were considered, for which
the exact solutions have been found [83]. Control of
nonlinear systems under Gaussian excitation has been
studied by approximate methods, such as stochastic aver-
aging[146], moment closure techniques[28]and EQL[12,
56,143145]. As EQL may be applied to relatively large
systems with geometrical as well as material nonlinearities,
combining EQL and linear optimal control theory opens an
avenue towards the control of relatively large nonlinear
structural systems. The general solution procedure is as
follows:
1. Start with a guess of the linearization coefficients.
2. Solve the Riccati equation for the linear control problem
and obtain the gains.
3. Calculate the expectations for the controlled linear
system.
4. Solve the linear system for the linearization coefficients.
5. Continue with step 2 until convergence.
A decoupling of the problem, i.e. the linearization of the
uncontrolled system followed by the calculation of theoptimal gains, would lead to different results, as the control
terms influence the expectations in the linearization
procedure. Recently, it has been shown [125]by means of
examples that the influence of the linearization criterion on
the optimal value of the cost functional seems to be
negligible.
2.6. Monte Carlo simulation
2.6.1. Introduction
For reasons mentioned in the beginning, the evaluation of
the stochastic response by MCS has particular advantages
over analytical approaches for nonlinear systems with largerstate-space dimension. First publications applying MCS in
stochastic structural mechanics appeared 30 years ago[119,
120]. The method has been since then applied and extended
by numerous authors[51,53,117,127]. In its simplest form,
denoted as direct Monte Carlo simulation (D-MCS), each of
the samples is generated independently according to a given
distribution. D-MCS is very robust and relatively easy to
apply. A sample size offew hundred independent realizations
is generally sufficient to obtain a suitable estimate for the
mean and variance and to provide information on the basic
shape of the distribution within the domain covered with
a probability of 1 2 10=N;whereNdenotes the sample size.However, it is not suited to provide information on the
tails of the distribution as often required for reliability
analysis. The reasons for this are quite obvious: dynamical
systems under service must be reliable, where possible
failures or malfunctions should be the exception to the rule;
i.e. failures should be rare, low probability events. This type
of inefficiency has been recognized since the early beginning
of Monte Carlo sampling at the end of the 1950s and
beginning of the 1960s, when so called variance reduction
methods were developed. In fact, books on Monte Carlo
sampling [45,54,112] are largely devoted to this subject,
addressing variance reduction procedures such as import-
ance sampling (IS), control variates and stratified sampling,
to mention the most common methods only. Other well
established approaches in Monte Carlo sampling, related to
nonlinear dynamical systems, are random walk concepts
used in physics, chemistry, economics and engineering[45].
Today, increasing the computational efficiency is the
main concern in research on MCS techniques. Two different
kinds of improvements are generally employed for an
increase in efficiency. The first kind improves the efficiency
in generating the samples. The otherso called variance
reduction techniquesincrease the efficiency by allowing
a reduced sample size for an acceptable variance of
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 115 7
-
8/10/2019 Proppe03_Equivalent Linearization and MC
8/15
the estimator. In the next sections, direct MCS is discussed
first, followed by a survey of methods to increase the
computational efficiency.
2.7. Direct Monte Carlo simulation
In D-MCS, one of the available random numbergenerators [58,59,84,96] is generally used to generate a
sequence of uniformly distributed pseudo-random numbers.
The random number generator must fulfill the statistical
tests of generating independent samples according to the
uniform distribution [46,85]. These pseudo-random num-
bers are then further translated according to the specified
distribution by using the inverse of the cumulative
distribution function or special algorithms like, e.g.
BoxMuller and polar Marsaglia [45,77] for specific
distributions. Since random number generators produce
statistically independent random numbers, correlation
between RVs must be constructed from independent RVs
by suitable transformations or algorithms. The generated
random numbers act as an input for the dynamical analysis,
which represents the computational most involved part.
For a sample sizeNthe variance, a point estimator pffor
the probability of failure pfhas the variance
var{pf}pf12pf=N: 16By accepting a toleranceewith the confidence level 1 2 d
Plpf2pfl , e $ 12 d; 17requires by employing Chebyshevs inequality a sample
size greater or equal to Nc
Ncpf12pf
de2 : 18
Since Chebyshevs inequality is valid for any sample size
and does not imply any restricting assumption, Nccustomarily specifies a larger sample size than necessary.
Assuming N large, a minimal sample size can be derived
from the central limit theorem, specified by the number
Nn
Nnpf12pfF2112 d=2
e
" #2; 19
where F denotes the standard normal cumulative
distribution function. Using Eq. (19) one should keep
in mind that Nn has to be sufficiently large to justify the
application of the central limit theorem.
Among all methods that utilize an N-point estimation in
the n-dimensional space, the Monte Carlo method has an
absolute estimation error that decreases with N21/2,
independently of the dimension n as seen from Eq. (16),
whereas all other approaches have errors which decrease
with N21=n at best[45].
As mentioned before, Monte Carlo sampling becomes
increasingly attractive as the dimension n of the problem
increases for n . 2: Hence MCS is often the only feasible
solution for problems with a larger dimension as one usually
encounters in technical applications. Besides being more
efficient than analytical-based approaches, is has the
advantages that the tools of deterministic analysis can be
fully exploited. Moreover, it is the most general applicable
tool available in stochastic mechanics, i.e. there is almost nostochastic problem where MCS cannot be applied.
Although MCS is more efficient than available alterna-
tives (if any), the required computational costs are often
considerably high. Naturally, the computational efforts
increase with the dimensionn, the required sample size N,
the complexity of the nonlinear structural models, and is
especially high when dynamic problems must be analyzed.
Hence, efficient techniques to compute the dynamical
response of the system for the input data should be employed
in order to reduce the computationally extensive tasks.
Moreover, so called variance reduction procedures[45,114]
can be used to further decrease the computational burden.
2.8. Increasing the efficiency of Monte Carlo simulation
2.8.1. Increasing the efficiency for generating samples
A considerable increase of computational efficiency is
often possible by exploiting the fact that the solutions in
different samples do not differ fundamentally, but just
gradually. In such cases, all (deterministic) solutions are
close to each other. Using a reference solution, the solutions
are in the neighborhood of the reference solution. Therefore,
the reference solution, which has to be determined only
once, is used to gain efficiency in the deterministic
computations. This procedure is described in the followingfor two examples, namely the Neumann series technique in
static stochastic structural analysis and the solution of
random eigenvalue problems in linear stochastic structural
dynamics. The Neumann expansion is a typical example for
such an approach, which has been proposed first in Ref.
[142]for static stochastic structural analysis in context with
stochastic finite elements and MCS. In static structural
analysis, the main computational part must be devoted to the
static solution of
ux Kx21fx; 20
whereu denotes the generalized displacement vector, ftheloading vector and K the stiffness matrix. All quantities
might depend on a set of RVs x. Since the factorization
(equivalent to the inversion) of the stiffness matrix requires
the main part of the solution, the Neumann expansion
represents the stiffness matrix by a reference matrixK0and
a deviator matrix DKxKx K0DKx; 21and factorizes only once the reference matrix K0.
The solution is then obtained by the sum
ux u0 2 u1u2 2 u3 ; 22
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 1158
-
8/10/2019 Proppe03_Equivalent Linearization and MC
9/15
with
u0x K210 fx; 23uix K210 DKxui21x fori . 0: 24The Neumann expansion is equivalent to placing the
deviator matrix on the right hand side
K0ux fx2 DKxux; 25and solving the above relation in an iteration which
converges for sufficiently large s
usx K210 {fx2 DKxus21x}; 26usings $ 0 and defining u21x 0:
It should be noted that the above iteration corresponds to
the well known class of NewtonRaphson procedures[11,
37] using the initial stress method. Newton Raphson
procedures and all its variants are applied in nonlinear
structural analysis since the development of finite element
analysis. The fact, that the Neumann expansion is just aspecial case of the NewtonRaphson schemes, indicates,
that the deterministic algorithm like NewtonRaphson can
be used to increase the computational efficiency of the
deterministic structural analyzes in context with MCS. It
seems that the power of available deterministic procedures
has not yet been fully exploited to treat uncertain/
deterministic nonlinear dynamical systems for determinis-
tic/stochastic loading by MCS.
Large linear FE-models subjected to dynamic loading are
usually analyzed by modal analysis. Within this procedure,
the computation of the modal properties like eigenfrequen-
cies and modes requires most of the computationalresources. In case of an uncertain or stochastic structural
system, it is not efficient to determine the modal properties
independently and anew for each sample. An increase in
efficiency can be obtained by using the fact that all solutions
of the random eigenvalue problem are closely related. This
circumstance has been exploited in Ref. [130], reducing
considerably the number of vector-iterations required to
obtain convergent eigensolutions. Other approaches pro-
posed[116]determine accurately a reference solution and
employ further component mode synthesis for a fast modal
update.
Alternatively, instead of solving the random eigenvalue
problem
Kxcix Mxcixlix; 27for each set of RVs x, a reference solution is determined
only once
K0ciM0cili: 28The effect of the deviator matrices for the stiffness,
damping, and mass, respectively
DKx Kx2K0; 29DCx Cx2C0;
DMx Mx2M0;are then treated on the right hand side of the equation of
motion, analogously to Eq. (26), leading with
ux Czx; 30
whereC comprises the set of eigenvectors, to the followingequation of motion in modal coordinatesz:
zi2jiwi_zi2w2izicTift2cTi DMxCz2c
Ti DCxC_z2cTi DKxCz: 31
The above system can be solved in an iterative manner,
where the modal solutions are slightly coupled by the right
hand side.
2.8.2. Importance sampling
For the case when the failure probability pfis very small,
standard Monte Carlo sampling is no longer efficient. This is
typically true when estimating failure probabilities where
pf, 1023:Suppose, for example, in a technical application,
the acceptable failure probability must be less than 1025.
Requiring a probability of 90% that the absolute sampling
error is less than 0.4 1025 would, then, according to Eq.
(18), require a sample size of at least 6.25 106
independent realizations. Such a huge sample size might,
in case the system is not exceptionally small, exceed
available resources even on powerful present day compu-
ters. Even though it might be possible to generate such large
sample sizes, using supercomputers and employing parallel
processing[68,69], the utilization of standard Monte Carlo
might be a waste of computational resources, since only asmall fraction of realizations of the sample contributes to the
desired statistical information.
Variance reduction procedures exploit additional a priory
information to reduce the necessary sample size N for a
specified accuracy. The widely used IS and controlled MCS
belong to this class of procedures and are discussed in this
and in Section 2.8.3.
IS gains its advantage by altering the density function
that generates the random samples. First publications
applying IS go back to Refs. [7072,74] using the first
computers. IS can be applied when an integral
pfR
nkFxdFx; 32
needs to be estimated by MCS, where Fx denotes thecumulative density function. In reliability problems the
kernelkxcan be interpreted as the indicator function Ixwhich assumes the value Ix 1 if x lies in the failuredomain which is usually expressed by the limit state
function gx , 0;and Ix 0 elsewhere. Introducing theimportance function Rx
Rx pxhx ; 33
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 115 9
-
8/10/2019 Proppe03_Equivalent Linearization and MC
10/15
as ratio between the original probability density function
and the IS density function hx; the above integral (32) isequivalent to
pfR
nkHxdHx; 34
whereHx is the IS distribution, andkHx kFxRx: 35IS can be used to reduce the variance of the kernelkxandtherefore the variance of the point estimate pf. Comparing
the variance of the two kernels, the following relation holds
var{kFx}2 var{kHx}
R
nk2Fx12RxdFx; 36
which guarantees a variance reduction if the right hand
side of Eq. (36) is positive. Choosing hx
. p
x
when
k2Fxpxis large and hx , px whenk2Fxpxis smallcontributes to the desired effect of reducing the variance
of the estimator, or equivalently reducing the sample size
N. The above relation also indicates that the IS density
should cover in reliability analysis the failure domain where
kFx Ix 1: Moreover, choosing hx kxpx=pfleads to zero variance thereby minimizing the variance
reduction. Since this optimal sampling density involves the
unknown estimator pf, it is merely of academic interest,
showing the existence of such a distribution.
The application of IS in nonlinear structural dynamics is
a relatively new development. On the contrary to the wide
application of IS in static structural mechanics, whichdeveloped in the last two decades to a well-established
methodology, IS seems presently to be in the early
development state. The obstacles in applying the IS concept
stems from the difficulty to establish a suitable IS density
hx for nonlinear structural problems. There are two mainsources for the problems to establish an IS density. One
difficulty stems from the large number of RVs involved to
represent the stochastic excitation. The other arises from the
nonlinear dynamic behavior which makes it hard to
associate a priori the excitation to the response. In stochastic
dynamics, the excitation is often modeled by white noise.
This implies a virtually infinite set of RVs. Since a very
large number of RVs cannot be handled efficiently using IS,
the stochastic excitation is usually expressed by stochastic
processes involving only a moderate number of RVs, say a
few hundred. Modeling the excitation either by a discrete
random pulse sequence[88]or employing the Karhunen
Loeve representation[49], it is usually possible to represent
the stochastic excitation process by a moderate number
of RVs.
IS is applied in context with reliability considerations
which are in stochastic structural dynamics first excursion
probabilities [110]. For linear systems, a quite efficient IS
density and methodology has been recently suggested in
Ref. [9]for solving the first excursion problem by IS. The IS
density is the union of all elementary failure regions Fikwhich are defined as the region in the RV space and which
cause a barrier crossing at instant tkdue to the ith output.
Using the unique linear relations between input and
response, the design points xpik; i.e. the nearest point to the
origin in standard normal space, can be established in astraightforward manner which define uniquely the elemen-
tary failure regions. Since the elementary failure probabil-
itiesPFikare known exactly, the optimal IS density for thecrossing problem is available. For computing the first
excursion probability, a weighted sum of the elementary
optimal IS densities is used, avoiding the inconvenience of a
multi-modal IS density which deviates considerably from
the optimal one.
IS can be extended to nonlinear systems with consider-
able computational effort to establish an IS density. For
cases where the design points of the elementary failure
events can be determined by nonlinear programming tools,
the above-described procedure can be applied to nonlinear
problems. The investigation in Ref. [39], for example,
utilizes the design pointsxpikof the elementary failure event,
to construct such an IS density. Since the search for the
elementary design point involves a constraint optimization
problem which requires the gradients, methods to determine
these gradients are necessary. Such gradients have been
derived in Ref. [87], applicable for the widely used
Newmark integration scheme. Further work following a
similar approach has been published inRef. [135]where a
hysteretic SDOF oscillator is investigated.
Since the IS density is usually rather difficult to estimate
for nonlinear dynamical systems, subset simulation has beensuggested in Ref. [8]. This work is an extension of an
adaptive importance sampling scheme [7] employing the
Metropolis algorithm, which generates a sequence of
samples that converges towards the optimal IS density.
Subset simulation introduces for a given failure event Fa
decreasing sequence of failure events F1 . F2 . F mF so that Fk >ki1Fi: Then, the final failure probability
pf PFm can be expressed as product of conditionalprobabilitiespf PF1
Qm21i1 PFi1lFi:The advantage of
this subset approach is that the failure probabilities PF1;PFi1lFi can be selected such that MC sampling is stilleffective, say in the order of 0.1.
An IS procedure applicable to Markov processes
described by Itos equation has been suggested in Ref.
[132]. This direction has been pursued further by other
authors[89,90,95]. The reported results show solutions for
restricted academic type problems, but the proposed
methods so far failed to analyze more realistic nonlinear
structural models.
2.8.3. Controlled Monte Carlo simulation
In recent years another concept applicable to dynamical
systemsdenoted as controlled MCShas been developed.
Contrary to IS, the methodology is self-adaptingand does not
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 11510
-
8/10/2019 Proppe03_Equivalent Linearization and MC
11/15
require detailed a priori information. It has no restrictions on
the type of structural model regarding linear, nonlinear
hysteretic and state-space dimension. Regarding efficiency,
IS can theoretically be much more efficient than any other
techniqueincluding controlled MCSif an efficient pro-
cedure is available to compute the IS density. However, one
should note that the efficiency of a procedure depends on howmuch information about a given problemthe method utilizes.
In essence, efficiency is generally gainedat the expense of the
loss of generality. As an example, direct MCS is very general
but not efficient in context with reliability problems. For
dynamic analyses, the IS density has been efficiently
determined for linear systems [9] and weakly nonlinear
systems only. Even for linear systems, the effort required to
determine the IS density is by far greater than the actual IS
simulation. The computational effort to estimate a suitable IS
density increases drastically for nonlinear dynamical sys-
tems. Moreover, for general nonlinear systems, the pro-
cedures established so far might even fail to converge to the
design point of the elementary failure events. If applicable at
all, the total amount of computational resources must be
considered. With respect to this total amount, an IS-
procedure cannot be considered as efficient, if the resources
to establish the IS density are orders of magnitude larger than
the IS-procedure itself. Hence, the following approach
should be seen as alternative, where the procedures
mentioned in Section 2.3.2 are not applicable, or cannot be
applied efficiently.
Controlled MCS methods have been developed for the
class of systems with Markovian properties and is designed
to increase the sampling density in the low probability
domain. This is made possible by modifying the weightsassociated with the realizations. In controlled MCS, the
weights are modified dynamically in an adaptive manner
contrary to IS where the weights (ratio between original and
IS density) are specified a priori.
Controlled MCS increases the sample density in the low
probability domain of the phase space and reduces the
sample density in the high probability domain. The increase
in the sample density is achieved by splitting[57,73,126]of
the sample. The technique of splitting is utilized
extensively in the reliability assessment of queues [5,57,
86]. The sample density is reduced by Russian Roulette[74,
103]or alternatively by a procedure denoted as clumping
[102].
Controlled MCS requires a selection criterion to
distinguish between important regions in the phase space
where the sample density needs to be increased and domains
where the sample density can be decreased. Since critical
dynamic responses are usually those which got an increased
energy input, first approaches employed an energy criterion
[101]increasing samples associated with a relatively high
mechanical energy (potential plus kinetic). This energy
criterion is not applicable for general stochastic dynamical
systems where energy is not defined. Moreover, for strongly
dissipative system, the criterion based on the mechanical
energy did not show a convincing performance. For this
reason, an alternative approach has been developed, for
which such energy criteria are no longer needed [55,103].
The key to this newly developed approach is to distribute
dynamically the sample density (realizations with different
weights or probabilities) approximately uniform within the
state space. By this approach, which does not require anyspecific selection criterion, the sample density is increased
automatically in the tails of the distribution.
In Ref. [55], the tendency to a uniform density of
realizations in the state space is achieved by clumping
subsequently the closest state pair to a single realization
until the number of realizations is reduced to half of the
sample size N. The closest pair is determined dynamically
by evaluating the smallest distance among all remaining
samples (including previously clumped realizations) in a
normalized state space. Subsequently, all realizations are
doubled. The approach suggested inRef. [103]also aims at
a uniform sampling distribution and employs Russian
Roulette and splitting. The importance of a sample is
measured by the distance to several closest neighbors which
allows to establish the selection criteria.
3. Conclusions
Due to its simplicity and computational efficiency, EQL
has become a standard tool of stochastic structural
dynamics, that allows for analyses of MDOF and FE-
systems with non-stationary excitation and hysteretic
material laws, respectively. Due to the introduction of
local coordinates, complex modal analysis, componentmode synthesis and the KarhunenLoeve representation,
EQL is now capable to deal with in principle arbitrary large
FE models.
There are however several points that need further
attention. First of all, despite some theoretical efforts, there
is still no easy to use method available for an estimate of the
approximation error. Next, the areas where EQL can be
applied are still to be fully discovered, e.g. a large part of the
nonlinear models available in commercial FE-codes and
employed by practicing engineers seems to be not suitable
for the computation of the linearization coefficients. The
applicability of EQL to practical problems is limited by the
fact that the method provides only second moment
information. In this regard, an extension of EQL by local
linearization techniques may give access to a wider range of
applications.
As EQL is generally not sufficiently accurate to
determine quantitatively the reliability of a nonlinear
structure under stochastic excitation, MCS procedures are
generally applied to estimate the reliability or its
complement the probability of failure for nonlinear
dynamical systems. Contrary to the static counterpart of
nonlinear systems, procedures to improve the efficiency
of MCS for the dynamic case are still in its early state of
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 115 11
-
8/10/2019 Proppe03_Equivalent Linearization and MC
12/15
development. Only recently, some progress has been
achieved in the fields of importance sampling and
controlled MCS, respectively. Except for linear or weakly
nonlinear systems, none of the available approaches is
satisfactory regarding the robustness of the approaches and
the computational efforts required. Hence, considerable
research effort might be necessary to arrive at satisfactoryprocedures to estimate the reliability of nonlinear dyna-
mical systems, especially of larger complex FE systems.
Acknowledgements
This section is sponsored by the IASSAR-CSMSE
Subcommittee 2 on Stochastic Dynamics (Y.K. Wen, chair).
References
[1] Ahmadi G, Tadjbakhsh I, Farshad M. On the response of nonlinear
plates to random loads. Acoustics 1978;40:31622.
[2] Apetaur M. Linearization of non-linear stochastically excited
dynamic systems by a modified second-order procedure. Vehicle
Syst Dyn 1986;15:11732.
[3] Apetaur M. Modified second order linearization procedure
problems encountered and their solution. Vehicle Syst Dyn 1988;
17:25565.
[4] Apetaur M, Opicka F. Linearization of non-linear stochastically
excited dynamic systems. J Sound Vib 1983;86(4):56385.
[5] Asmussen S, Rubinstein R. Steady-state rare events simulation in
queuing models and its complexity properties. In: DhashlowJ, editor.
Advancesin queueing. Boca Raton, FL: CRCPress; 1995.p. 79102.
[6] Atalik TS, Utku S. Stochastic linearization of multi-degree-of-
freedom non-linear systems. Earthquake Engng Struct Dyn 1976;4:41120.
[7] Au SK, Beck JL. A new adaptive importance sampling scheme.
Struct Safety 1999;21:13558.
[8] Au SK, Beck JL. Subset simulation: a new approach to calculating
small failure probabilities. Proceedings of the International Con-
ference on Monte Carlo Simulation, Monte Carlo, Monaco; 2000.
[9] Au SK, Beck JL. First excursion probabilities for linear systems by
very efficient importance sampling. Probab Engng Mech 2001;6:
193207.
[10] Baber TT, Noori MN. Random vibration of pinching hysteretic
systems. J Engng Mech (ASCE) 1984;110(7):103649.
[11] Bathe K-J. Finite element procedures. Englewood Cliffs, NJ:
Prentice Hall; 1996.
[12] Beaman JJ. Non-linear quadratic Gaussian control. Int J Control
1984;39(2):34361.[13] Beaman JJ, Hedrick JK. Improved statistical linearization for
analysis and control of nonlinear stochastic systems. Part I. An
extended statistical linearization technique. Trans ASME, J Dyn
Syst, Meas, Control 1981;103:1421.
[14] Bergman LA, Spencer Jr.BF. In: Bellomo N, Casciati F, editors.
Robust numerical solution of the transient FokkerPlanck equation
for nonlinear dynamical systems. Proceedings of the IUTAM
Symposium on Nonlinear Stochastic Mechanics, Turin, Italy, Berlin:
Springer; 1991. p. 4960.
[15] Bernard P. Stochastic linearization: what is available and what is not.
Comput Struct 1998;67:918.
[16] Bernard P, Taazount M. Random dynamics of structures with gaps:
simulation and spectral linearization. Nonlinear Dyn 1994;5:
31335.
[17] Bernard P, Wu L. Stochastic linearization: the theory. J Appl Probab
1998;35:71830.
[18] Bruckner A, Lin YK. Generalization of the equivalent linearization
method for non-linear random vibration problems. Int J Non-Linear
Mech 1987;22(3):22735.
[19] Busby Jr. HR, Weingarten VI. Response of a nonlinear beam to
random excitation. J Engng Mech Div 1973;99(EM1):5568.
[20] Casciati F. Stochastic dynamics of hysteretic media. Struct Safety
1989;6:25969.
[21] Casciati F, Faravelli L. Stochastic finite element analysis of
nonlinear media. Comput Exp 1989;176:1337.
[22] Casciati F, Faravelli L. Fragility analysis of complex structural
systems. Taunton: Research Studies Press; 1991.
[23] Casciati F, Faravelli L, Hasofer AM.A newphilosophy forstochastic
equivalent linearization. Probab Engng Mech 1993;8:17985.
[24] Caughey TK. Response of a nonlinear string to random loading.
Trans ASME, J Appl Mech 1959;26:3413.
[25] Caughey TK. Response of Van der Pols oscillator to random
excitation. Trans ASME, J Appl Mech 1959;26:3458.
[26] Caughey TK. Random excitation of a system with bilinear
hysteresis. Trans ASME, J Appl Mech 1960;27:64952.
[27] Caughey TK. Equivalent linearization techniques. J Acoust Soc Am
1963;35:170611.[28] Chang RJ. Optimal linear feedback control for a class of nonlinear
nonquadratic non-Gaussian problems. Trans ASME, J Dyn, Meas
Control 1991;113(4):56874.
[29] Chang RJ. Non-Gaussian linearization method for stochastic
parametrically and externally excited nonlinear systems. Trans
ASME, J Dyn Syst, Meas, Control 1992;114:206.
[30] Chang T-P, Mochio T, Samaras E. Seismic response analysis of
nonlinear structures. Probab Engng Mech 1986;1(3):15766.
[31] Chang RJ, Young GE. Methods and Gaussian criterion for statistical
linearization of stochastic parametrically and externally excited
nonlinear systems. Trans ASME, J Appl Mech 1989;56:17985.
[32] Craig Jr RR, Chang C-J. On the use of attachment modes in
substructure coupling for dynamic analysis. AIAA Paper 77-405 and
AIAA/ASME 18th Structures, Structural Dynamics and Material
Conference, San Diego, CA; 1977. p. 8999.[33] Craig Jr. RR. Structural dynamics: an introduction to computer
methods. New York: Wiley; 1981.
[34] Crandall SH. Heuristic and equivalent linearization techniques for
random vibration of nonlinear oscillators. VIII International
Conference Nonlinear Oscillations, Prague; 1979. p. 21126.
[35] Crandall SH. In: Ramnath RV, Hedrick JK, Paynter HM, editors. On
statistical linearization for nonlinear oscillations. Nonlinear system
analysis and synthesis: techniques and applications, vol. 2. New
York: ASME; 1980. chapter 12.
[36] Crandall SH. Is stochastic equivalent linearization a subtle flawed
prcoedure? Probab Engng Mech 2001;16(2):16976.
[37] Crisfield MA, Non-linear finite element analysis of solids and
structures, vol. 1. West Sussex: Wiley; 1991.
[38] Dobson S, Noori M, Hou Z, Dimentberg M, Baber T. Modeling and
random vibration analysis of SDOF systems with asymmetrichysteresis. Int J Non-Linear Mech 1997;32(4):66980.
[39] Der Kiureghian A, Li C-C. In: Lin YK, Su TC, editors. Statistics of
fractional occupation time for nonlinear stochastic response.
Proceedings of the 11th Engineering Mechanics Conference, Fort
Lauderdale: ASCE; 1996. p. 1169.
[40] Elishakoff I, Colajanni P. Stochastic linearization critically re-
examined. Chaos, Solitons Fractals 1997;8(12):195772.
[41] Elishakoff I, Zhang R. In: Bellomo N, Casciati F, editors.
Comparison of the new energy-based versions of the stochastic
linearization technique. Proceedings IUTAM Symposium Turin,
Berlin: Springer; 1992. p. 20112.
[42] Emam H, Pradlwarter HJ, Schueller GI. A computational procedure
for the implementation of equivalent linearization in finite element
analysis. Earthquake Engng Struct Dyn 2000;29:117.
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 11512
-
8/10/2019 Proppe03_Equivalent Linearization and MC
13/15
[43] Falsone G. Stochastic linearization of MDOF systems under
parametric excitation. Int J Non-Linear Mech 1992;27(6):102537.
[44] Falsone G, Vasta M. On the approximated solution of non-linear
systems under non-Gaussian excitations. Proceedings of the Sixth
Specialty Conference on Probabilistic Mechanics and Structural and
Geotechnical Reliability, Denver, Colorado; July 8 10, 1992. p.
1403.
[45] Fishman GS. Monte Carlo: concepts algorithms, and applications.
Springer series in operations research, New York: Springer; 1996.
[46] Fishman GS, Moore LR. An exhaustive analysis of multiplicative
congruential randomnumber generators with modulus 231 2 1 .J A m
Stat Assoc 1982;77:12936.
[47] Foliente GC, Singh MP, Noori MN. Equivalent linearization of
generally pinching hysteretic, degrading systems. Earthquake Engng
Struct Dyn 1996;25:61129.
[48] Foster ET. Semi-linear random vibrations in discrete systems. Trans
ASME, J Appl Mech 1968;35(3):5604.
[49] Ghanem R, Spanos P. Stochastic finite elements: a spectral approach.
New York: Springer; 1991.
[50] Goto H, Iemura H. Linearization techniques for earthquake response
of simple hysteretic structures. Proc Jpn Soc Civil Engng 1973;212:
10919.
[51] Grigoriu M. Simulation of nonstationary Gaussian processes byrandom trigonometric polynomials. ASCE J Engng Mech 1993;
119(2):32843.
[52] Grigoriu M. Equivalent linearization for Poisson white noise input.
Probab Engng Mech 1995;10:4551.
[53] Grigoriu M. Simulation of stationary non-Gaussian translation
processes. ASCE J Engng Mech 1998;124(2):1216.
[54] Hammersley JM, Handscomb JC. Monte Carlo methods. London:
Methuen; 1964.
[55] Harnpornchai N, Pradlwarter HJ, Schueller GI. Stochastic analysis of
dynamical systems by phase-space-controlled Monte Carlo simu-
lation. Comput Meth Appl Mech Engng 1999;168:27383.
[56] Heess G. Application of state-space statistical linearization to
optimal stochastic control of non-linear systems. Int J Control
1970;11(4):697701.
[57] Heidelberger P. Fast simulation of rare events in queuing andreliability models. ACM Trans Modeling Comput Simul 1995;5:
4385.
[58] Hellekalek P. Good random number generators are (not so) easy to
find. Math Comput Simul 1998;46:485505.
[59] Hennecke, M. Random number generator homepage; 2000,http://
www.uni-karlsruhe.de/, RNG/:.
[60] Hsu CS. Cell-to-cell mapping: a method of global analysis for
nonlinear systems. New York: Springer; 1987.
[61] Hsu CS. In: Kliemann W, Namachchivaya SN, editors. Domain-to-
domain evolution by cell mapping, in nonlinear dynamics and
stochastic mechanics. Boca Raton, FL: CRC Press; 1995. p. 4568.
[62] Hurtado JE, Barbat AH. Improved stochastic linearization method
using mixed distributions. Struct Safety 1996;18(1):4962.
[63] Iwan WD, Krousgrill CM. Equivalent linearization for continuous
dynamical systems. Trans ASME, J Appl Mech 1980;50:41520.[64] Iwan WD, Lutes LD. Response of the bilinear hysteretic system to
stationary random excitation. J Acoust Soc Am 1968;43:54552.
[65] Iwan WD, Whirley RG. Nonstationary equivalent linearization of
nonlinear continuous systems. Probab Engng Mech 1993;8:27380.
[66] Iwankiewicz R, Nielsen SRK. Dynamic response of hysteretic
systems to Poisson distributed pulse trains. Probab Engng Mech
1992;7:13548.
[67] Iyengar RN. Higher order linearization in non-linear random
vibration. Int J Non-Linear Mech 1988;23(5):38591.
[68] Johnson EA, Wojtkiewicz SF, Bergman LA. In: Spanos PD, editor.
Some experiments with massively parallel computation for Monte
Carlo simulation of stochastic dynamical systems. Computational
stochastic dynamics, Rotterdam: Balkema; 1995. p. 32536. ISBN
90 5410527 5.
[69] Johnson EA, Woitkiewicz SF, Bergman LA, Spencer Jr. BF.
Observations with regard to massively parallel computation for
Monte Carlo simulation of stochastic dynamical systems. Int J Non-
linear Mech 1997;32(4):72134.
[70] Kahn H. In: Hurd CC, editor. Modification of the Monte Carlo
method. Seminar on scientific computations, Nov 16 18, 1949,
IBM; 1950.
[71] Kahn H. Random sampling (Monte Carlo) techniques in neutron
attenuation problems-I. Nucleonics 1950;6:2733. see also page 37.
[72] Kahn H. Random sampling (Monte Carlo) techniques in neutron
attenuation problems-II. Nucleonics 1950;6:605.
[73] Kahn H, Harris TE. Estimation of particle transmission by random
sampling. Nat Bur Stand Appl Math Ser 1951;12:2730.
[74] Kahn H. In: Mayer MA, editor. Use of different Monte Carlo
sampling techniques. Symposium on Monte Carlo methods, New
York: Wiley; 1956. p. 14690.
[75] Kazakov IE. Approximate probabilistic analysis of the accuracy of
operation of essentially nonlinear systems. Automat Remote Control
1956;17:42350.
[76] Kazakov IE. Statistical analysis of systems with multi-dimensional
non-linearities. Automat Remote Control 1965;26:45864.
[77] Kloeden PE, Platen E, Schurz H. Numerical solution of SDE through
computer experiments. Berlin: Springer; 1991.[78] Kobori T, Minai R, Suzuki Y. Statistical linearization techniques for
hysteretic structures with earthquake excitations. Bull Disaster
Prevent Inst, Kyoto Univ 1973;23(215):11135.
[79] Kolovskii MZ. Estimating the accuracy of solutions obtained by the
method of statistical linearization. Automat Remote Control 1966;
27:1692701.
[80] Kottalam J, Lindenberg K, West BJ. Statistical replacement for
systems with delta-correlated fluctuations. J Stat Phys 1986;42(5/6):
9791009.
[81] Kozin F. In: Ziegler F, Schueller GI, editors. The method of
statistical linearization for non-linear stochastic vibrations. Proceed-
ings IUTAM Symposium on Nonlinear Stochastic Dynamic
Engineering Systems, Berlin: Springer; 1988. p. 4556.
[82] Krylov N, Bogoliubov N. Introduction to non-linear mechanics.
Princeton, NJ: Princeton University Press; 1947.[83] Kwakernaak H, Sivan R. Linear optimal control systems. New York:
Wiley; 1972.
[84] LEcuyer P. In: Jerry Banks, editor. Handbook on simulation. New
York: Wiley; 1997.
[85] LEcuyer P. In: Swain JJ, Craig RC, Wilson JR, Goldsman D,
editors. Testing random number generators. Proceedings of the 1992
Winter Simulation Conference, Arlington, VA, Piscataway, NJ:
IEEE Press; 1992. p. 30513.
[86] Lewis EE, Bohm F. Monte Carlo simulation of Markov unreliability
models. Nucl Engng Des 1984;77:4962.
[87] Li C-C, Der Kiureghian A. In: Lemaire M, Fabre JL, Mebarki A,
editors. Mean outcrossing rate of nonlinear response to stochastic
input. Application of statistics and probability, Rotterdam: Balkema;
1995. p. 11121.
[88] Lin YK. Probabilistic theory of structural dynamics. Huntington,New York: Robert E. Krieger Publication; 1976.
[89] Macke M. In: Melchers RE, Stewart MG, editors. Variance reduction
in Monte Carlo simulation of dynamical systems. Applications of
statistics and probability, Rotterdam: Balkema; 2000.
[90] Macke M, Harnpornchai N. In: Corotis R, Schueller GI, Shinozuka
M, editors. Importance sampling of dynamic systems: a comparative
study. Proceedings of the Eighth International Conference on
Structural Safety and Reliability, CD-ROM; 2001.
[91] Bingqi M. Direct integration variance prediction of random response
of nonlinear systems. Comput Struct 1993;46(6):97983.
[92] Micaletti RC, Cakmak AS, Nielsen SRK, Koyluoglu HU. Error
analysis of statistical linearization with Gaussian closure for
large-degree-of-freedom systems. Probab Engng Mech 1998;
13(2):7784.
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 115 13
http://www.uni-karlsruhe.de/RNG/http://www.uni-karlsruhe.de/RNG/http://www.uni-karlsruhe.de/RNG/http://www.uni-karlsruhe.de/RNG/http://www.uni-karlsruhe.de/RNG/http://www.uni-karlsruhe.de/RNG/http://www.uni-karlsruhe.de/RNG/http://www.uni-karlsruhe.de/RNG/ -
8/10/2019 Proppe03_Equivalent Linearization and MC
14/15
[93] Mochio T, Samaras E, Shinozuka M. Stochastic equivalent
linearization for finite element-based reliability analysis. Proc
Fourth Int Conf Struct Safety Reliab 1985;37584.
[94] Naess A, Johnsen JM. Response statistics of nonlinear, compliant
offshore structures by the path integral solution method. Probab
Engng Mech 1993;8(2):91106.
[95] Naess A, Skaug C. In: Fryba P, Naprstek J, editors. Estimates of
failure probability by importance sampling for dynamic systems.
Structural dyanmics-EURODYN99, Rotterdam: Balkema; 1999.
p. 27782.
[96] Niederreiter H. Random number generation and quasi-Monte
Carlo methods. CBMSNSF Regional Conference Series in
Applied Mathematics, Society for Industrial and Applied
Mathematics; 1992.
[97] Ohtori Y, Spencer BF. Semi-implicit integration algorithm for
solution of nonlinear stochastic vibration problems. Eighth ASCE
Speciality Conference on Probabilistic Mechanics and Structural
Reliability 2000;CD-ROM:PMC2000-323.
[98] Park YJ, Wen YK, Ang AHS. Random vibration of hysteretic systems
under bi-directional ground motion. Earthquake Engng Struct Dyn
1986;14:54357.
[99] Polidori DC, Beck JL. Approximate solutions for non-linear random
vibration problems. Probab Engng Mech 1996;11:17985.[100] Pradlwarter HJ, Li W-L. On the computation of the stochastic
response of highly nonlinear large MDOF-systems modeled by finite
elements. Probab Engng Mech 1991;6(2):10916.
[101] Pradlwarter HJ, Schueller GI, Melnik-Melnikov PM. Reliability of
MDOF-systems. Probab Engng Mech 1994;9(4):23543.
[102] Pradlwarter HJ, Schueller GI. On advanced Monte Carlo simulation
procedures in stochastic structural dynamics. Int J Non-Linear Mech
1997;32(4):73544.
[103] Pradlwarter HJ, Schueller GI. Assessment of low probability events
of dynamical systems by controlled Monte Carlo simulation. Probab
Engng Mech 1999;14:21327.
[104] Pradlwarter HJ. Stochastic non-linear response distribution by local
statistical linearization. Int J Non-Linear Mech 2001;36(7):
113551.
[105] Pradlwarter HJ. Deterministic integration algorithms for stochasticresponse computations of large FE-systems. Comput Struct 2002;80:
1489502.
[106] Pradlwarter HJ, Schueller GI, Schenk CA. A Computational
procedure to estimate the stochastic dynamic response of large
non-linear FE-models. Comput Meth Appl Mech Engng 2001; in
press.
[107] Proppe C. Stochastic linearization of dynamical systems under
parametric Poisson white noise excitation. Int J Non-Linear Mech
2002; 2003;18(4):54355.
[108] Proppe C, Schueller GI. Equivalent Linearization Revisited. Ko JM,
Xu YL, editors. Proc Int Conf Adv Struct Dyn, Hong Kong
December 13 15, 2000;1207 14.
[109] Roberts JB. The response of an oscillator with bilinear hysteresis to
random excitation. Trans ASME, J Appl Mech 1978;45:9238.
[110] Roberts JB. First passage time for randomly excited non-linearoscillators. J Sound Vib 1986;109(1):3350.
[111] Roberts JB, Spanos PD. Random vibration and statistical lineariza-
tion. Chichester: Wiley; 1990.
[112] Rubinstein RY. Simulation and the Monte Carlo method. New York:
Wiley; 1981.
[113] Schueller GI, Pradlwarter HJ, Vasta M, Harpornchai N. In: Shiraishi
N, Shinozuka M, Wen YK, editors. Benchmark study on non- linear
stochastic structural dynamics. Proceedings of the Seventh Inter-
national Conference on Structural Safety and Reliability, Rotterdam:
Balkema; 1998.
[114] Schueller GI, Stix R. A critical appraisal of methods to determine
failure probabilities. J Struct Safety 1987;4:293309.
[115] Schueller GI, Pradlwarter HJ. On the stochastic response of
nonlinear FE-models. Arch Appl Mech 1999;69:76584.
[116] Schueller GI, Pradlwarter HJ, Szekely GS. On the random
eigenvalue problem. In: Topping BHV, editor. Computational
mechanics: techniques and developments. Proceedings of the Fifth
International Conference on Computational Structures Technology,
Leuven, Belgium, Edinburgh: Saxe-Coburg Publication; 2000. p.
17380.
[117] Schueller GI, Spanos PD, editors. Monte Carlo simulation.
Proceedings of the International Conference on Monte Carlo
Simulation, Monte Carlo, Lisse: A.A. Balkema; 2000.
[118] Seide P. Nonlinear stresses and deflections of beams subjected to
random time dependent uniform pressure. Trans ASME, J Engng Ind
1976;98:101420.
[119] Shinozuka M. Monte Carlo solution of structural dynamics. Int J
Comput Struct 1972;2:85574.
[120] Shinozuka M, Wen Y-K. Monte Carlo solution of nonlinear
vibrations. AIAA J 1972;10(1):3740.
[121] Simulescu I, Mochio T, Shinozuka M, Samaras M. Stochastic
equivalent linearization method in nonlinear FEM. J Engng Mech
(ASCE) 1989;115(3):47592.
[122] Skrzypczyk J. Accuracy analysis of statistical linearization methods
applied to nonlinear dynamical systems. Rep Math Phys 1995;36(1):
120.
[123] Socha L. Probability density equivalent linearization technique fornonlinear oscillator with stochastic excitations. Z AngewMath Mech
2000; in press.
[124] Socha L, Pawleta M. Corrected equivalent linearization of stochastic
dynamic systems. Mach Dyn Prob 1994;7:14961.
[125] Socha L, Proppe C. Optimal control of the Duffing oscillator under
non-Gaussian external excitation. Eur J Mech A 2002; in press.
[126] Spanier J, Gelbard EM. Monte Carlo principles and neutron transport
problems. Addison-Wesley series in computer science and infor-
mation processing, Reading, MA: Addison-Wesley; 1969.
[127] Spanos PD, Zeldin BA. Monte Carlo treatment of random fields: a
broad perspective. Appl Mech Rev 1998;51(3):21937.
[128] Spencer Jr. BF, Bergman LA. On the numerical solution of the
Fokker Planck equation for nonlinear stochastic systems. Nonlinear
Dyn 1993;4:35772.
[129] Sperling L. Analytische Naherungsmethoden zur Untersuchungnichtlinearer, stochastisch erregter Systeme. Habilitationsschrift,
Technische Hochschule Otto von Guericke, Magdeburg; 1979.
[130] Szekely GS, Schueller GI. Computational procedure for a fast
calculation of eigenvectors and eigenvalues of structures with
random properties. Comput Meth Appl Mech Engng 2001;8(810):
799816.
[131] Takemiya H, Lutes LD. Stationary random vibration of hysteretic
systems. Trans ASCE, J Engng Mech Div 1977;103(EM4):
67388.
[132] Tanaka H. Application of an importance sampling method to time-
dependent reliability analyses using the Girsanov transformation. In:
Shiraishi N, Shinozuka M, Wen YK, editors. Structural safety and
reliability. Rotterdam: Balkema; 1998. p. 4118.
[133] To CWS. Direct integration operators and their stability for random
response of multi-degree-of-freedom systems. Comput Struct 1988;30(4):86574.
[134] Tylikowski A, Marowski W. Vibration of a non-linear single degree
of freedom system due to Poissonian impulse excitation. Int J Non-
Linear Mech 1986;21(3):22938.
[135] Vijalapura PK, Conte JP, Meghella M. In: Melchers RE, Stewart
MG, editors. Time-variant reliability analysis of hysteretic SDOF
systems with uncertain parameters and subjected to stochastic
loading. Application of statistics and probability, Rotterdam:
Balkema; 2000.
[136] Wehner MF, Wolfer WG. Numerical evaluation of path-integral
solutions to Fokker Planck equations. Phys Rev A 1983;27(5):
266370.
[137] Wen YK. Equivalent linearization for hysteretic systems under
random excitation. Trans ASME, J Appl Mech 1980;47:1504.
C. Proppe et al. / Probabilistic Engineering Mechanics 18 (2003) 11514
-
8/10/2019 Proppe03_Equivalent Linearization and MC
15/15
[138] Wen YK. Methods of random vibration for inelastic structures. Appl
Mech Rev 1989;42(2):3952.
[139] Wen YK, Yeh CH. Bi-axial and torsional response of inelastic
structures under random excitation. Proceedings of the Symposium
on Stochastic Structural Dynamics. University of Illinois; 1988.
[140] Wojtkiewicz SF, Bergman LA. Numerical solution of a four-
dimensional Fokker-Planck equation. Bathe KJ, editor. Comput
Fluid Solid Mech 2001;16735.
[141] Wu WF. Comparison of Gaussian closure technique and equivalent
linearization method. Probab Engng Mech 1987;2(1):28.
[142] Yamazaki F, Shinozuka M, Dasgupta G. Neumann expansion for
stochastic finite element analysis. J Engng Mech, ASCE 1988;
114(8):133554.
[143] Yang JN, Li Z, Vongchavalitjul S. Generalization of optimal control
theory: linear and nonlinear control. J Engng Mech Div (ASCE)
1984;120(2):26683.
[144] Yoshida K. A method of optimal control of non-linear stochastic
systems with non-quadratic criteria. Int J Control 1984;39(2):
27991.
[145] Young JJ, Chang RJ. Optimal control of stochastic pa