Multivariate probability distribution of ordered peaks of vector

Gaussian random processes

C.S. Manohar
a,
*
,1
, Roger Ghanem
b
a
Department of Civil Engineering, Indian Institute of Science, Bangalore 560 012, India
b
Department of Civil Engineering, Johns Hopkins University, Baltimore, MD 21218, USA
Received 21 August 2003; revised 24 May 2004; accepted 6 July 2004
Abstract
The problem of determining the joint probability distribution of ordered peaks of jointly stationary Gaussian random processes is
considered. The solution is obtained by modeling the number of times a specied threshold is crossed by the component processes as a
multivariate Poisson process. Based on this, the joint probability distribution of the time required for the nth crossing of a specied level with
positive slope is derived. This formulation is further extended to derive the joint distribution of ordered peaks in a given time interval. An
illustrative example on a bivariate Gaussian random process is presented and the analytical predictions are shown to compare reasonably well
with corresponding results from Monte Carlo simulations. Also presented is an analysis of response of a randomly driven multi-degree of
freedom system with emphasis on the sensitivity of ordered peak characteristics with respect to changes in system parameters. It is
demonstrated that higher order statistics are generally more sensitive to changes in system characteristicsa property that has potential for
application in structural model updating and damage detection.
q 2004 Elsevier Ltd. All rights reserved.
Keywords: Probability distribution; Gaussian random process; Ordered peaks
1. Introduction
Extreme values of structural capacities and loads play
decisive role in the study of reliability of structural systems
[14]. The development of response spectrum-based methods
in earthquake engineering and gust factor method in wind
engineering is strongly based on the theory of extremes of
random processes [15]. The study of extreme values of a
sequence of random variables forms a subset of a more
broad-based study on order statistics. The properties of
order statistics in general, and, extreme values in particular,
of sequence of random variables have been widely studied
in the existing literature by mathematicians and engineers.
From a mathematical perspective, the books by Gumbel
[7], Galambos [6], Resnick [17], Castillo [4] and Kotz and
Nadarajah [13] provide comprehensive overviews on the
subject. A few highlights of established results, as can be
deduced from a study of these books, can be summarized
briey as follows. For a sequence of identical and
independently distributed (IID) random variables, there
exist fundamentally three forms of asymptotic distributions
for the extremes. Methods to establish domains of attraction
of these asymptotic distributions are available. The tail
behavior of the underlying random variables forms the basis
on which these domains are established. Similarly, it is
possible to identify feasible limit distributions for kth order
statistics in the IID case, with separate analysis for upper or
lower, moderately upper and lower and central order
statistics. Results on extremes of sequence of identical,
but dependent, random variables include, conditions on
dependence structure, which ensures the continued validity
of limiting distributions that are strictly valid only for IID
case and results on more general forms of dependence
sequences. The treatment of asymptotic forms of extremes
of nonstationary and dependent sequence of random
0266-8920/$ - see front matter q 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.probengmech.2004.07.002
Probabilistic Engineering Mechanics 20 (2005) 8796
www.elsevier.com/locate/probengmech
* Corresponding author. Tel.: C91 80 2293 3121; fax: C91 80 2360
0404.
E-mail address: [email protected] (C.S. Manohar).
1
Visiting Associate Professor, Johns Hopkins University, Baltimore, MD
21218, USA.
variables is rather limited [11]. Studies on asymptotic forms
of multivariate extreme value distributions of vector
sequences of random variables have also been conducted.
Here results are available mainly for distinct IID sequences
such as U
i
and V
i
(iZ1,2,.,n), where the random variables
U
i
and V
j
are mutually independent except when iZj.
Results on asymptotic joint distributions of ordered
statistics of U
i
and V
i,
(iZ1,2,.,n) are also available.
In structural engineering applications, the study of
ordered peaks of random processes has received limited
attention. These studies are of interest, for example, in the
study of progressive failure of ductile structures subjected to
dynamic loads such as earthquake ground accelerations.
Here, the exceedance of the structural response across a
specied level only once over specied time duration may
not signal the failure of the system. Consequently, proper-
ties of lower order peaks become important, especially, if
attention is focused on response levels that make mild
excursions into nonlinear regimes but nevertheless remain
higher than permissible linear limits. The work by Amini
and Trifunac [1], Gupta and Trifunac [8,9], Basu and Gupta
and Basu et al. [2,3] contain discussions on related issues.
The starting point in these studies is the model for the PDF
of local maxima of Gaussian random processes [16]. Amini
and Trifunac derive the probability distribution function
(PDF) of the nth order peak by assuming that the peaks are
independent, and, also, that the probability of exceeding a
specied level by all the peaks is the same as that by the rst
order peak. Gupta and Trifunac [8] relax the second of these
assumptions. Basu et al. [2,3] have employed simulation-
based method and Markovian modeling procedures to
investigate the properties of ordered peaks by relaxing the
assumption of independence of peaks. The note by Gupta
and Trifunac [8] reports on numerical experiments aimed at
evaluation of relative performance of procedures due to
Amini and Trifunac [1], Gupta and Trifunac [9] and Basu
et al. [3]. This study concludes that the assumption of
independence of peaks in deriving the PDF of ordered peaks
is not restrictive.
In the present study, we develop a model for joint PDF of
vector of ordered peaks associated with a vector stationary
Gaussian random process. This distribution function is
shown to be related to the joint distribution of time for
crossing of a given threshold for the nth time. This, in turn,
is deduced in terms of a multivariate counting process for
modeling the number of times a specied threshold is
crossed. This work is in continuation of a recent study by
Gupta and Manohar [10] in which multivariate distributions
for extremes of vector Gaussian random process have been
developed and applied to problems of time variant structural
system reliability analysis. We foresee that the properties of
ordered peaks of random structural responses could serve as
useful tools for structural model updating and damage
detection. The present study offers preliminary evidence in
support of this prognosis.
2. Analysis
Consider an n!1 vector random process X(t), t2T,
whose elements are made up of zero mean, jointly
stationary, Gaussian random processes. Let R(t) and
S(u) denote, respectively, the covariance matrix and the
power spectral density (PSD) function matrix of X(t). The
elements of X(t) are assumed to be differentiable at least
once. In this study, we seek to obtain the following
descriptors of X(t):
1. N
i
(a
i
,0,T)Zthe number of times the random process
X
i
(t) crosses the level a
i
with positive slope in the time
interval 0T. For a given T, NZ{N
i
(a
i
,0,T} constitutes
a vector of discrete random variables. What is the
multivariate PDF of the vector N?
2. T
ij
Zthe time taken by X
i
(t) to cross a threshold a
i
with
positive slope for the jth time. What is the joint PDF of
T
ij
for iZ1,2,.,k%n and jZ1,2,.?
3. X
ij
Zthe jth highest peak of X
i
(t) over a duration 0T.
What is the joint PDF of X
ij
for iZ1,2,.,k%n and
jZ1,2,.?
It may be remarked here that for the special case of nZ1,
N can be modeled approximately as a Poisson random
variable from which it can be shown that T
11
is an
exponential random variable and X
11
is a Gumbel random
variable [16]. In a recent study Gupta and Manohar [10]
have considered the case of nO1 and has developed
multivariate Gumbel models for X
i1
(iZ1,2,.,k%n).
2.1. Level crossings
To arrive at the joint PDF of N
i
(a
i
,0,T), iZ1,2,.,k%n,
for a given T, we assume that these counting processes
constitute a vector of multi-variate Poisson random vari-
ables. To clarify this, we consider the case of kZ2. We
introduce the transformation
N
1
(a
1
; 0; T) = U
1
- U
3
N
2
(a
2
; 0; T) = U
2
- U
3
(1)
where U
i
(iZ1,2,3) are three mutually independent Poisson
random variables with parameters l
i
(iZ1,2,3), respecti-
vely. These parameters are as yet unknowns. It can be
shown that N
1
(a
1
,0,T) and N
2
(a
2
,0,T) are Poisson random
variables with parameters (l
1
Cl
3
) and (l
2
Cl
3
), respecti-
vely. It may be noted that this construct of multivariate
Poisson random variables has been discussed in the existing
literature [12]. It can also be shown that the covariance of
N
1
(a
1
,0,T) and N
2
(a
2
,0,T) is equal tol
3
. Based on this we
get the equation
1 0 1
0 1 1
0 0 1
_

_
_

_
l
1
l
2
l
3
_

_
_

_
Z
(N
1
(a
1
; 0; T))
(N
2
(a
2
; 0; T))
Cov[N
1
(a
1
; 0; T); N
2
(a
2
; 0; T)]
_

_
_

_
(2)
C.S. Manohar, R. Ghanem / Probabilistic Engineering Mechanics 20 (2005) 8796 88
Here (.) denotes the mathematical expectation operator
and Cov[.] denotes the covariance. The counting
processes N
i
(a
I
,0,T) (IZ1,2), are well known to be
related to the parent processes X
i
(t) (IZ1,2), through the
relations [16]
N
i
(a
i
; 0; T) Z
_
T
0
d[X
i
(t)Ka
i
]U[
_
X
i
(t)]
_
X
i
(t)dt (3)
Here U(.) and d(.) are, respectively, the Heaveside and
Diracs delta functions. Using the above relation, it is
possible to calculate the quantities appearing on the right
hand side of Eq. (2). The determination of covariance
function appearing in Eq. (2) requires the evaluation of a
six dimensional integral. Gupta and Manohar [10] have
employed symbolic manipulation tools to reduce this
integral to a two-dimensional integral that requires
numerical evaluation. The unknowns l
i
(iZ1,2,3) can
thus be evaluated once the right hand side in Eq. (2) is
determined. To construct the joint PDF of N
1
(a
1
,0,T) and
N
2
(a
2
,0,T) we rst construct the joint characteristic
function
F
12
(u
1
; u
2
) Z(exp[iu
1
N
1
Ciu
2
N
2
])
Z(exp[iu
1
(U
1
CU
3
)Ciu
2
(U
2
CU
3
)])
Z(exp[iu
1
U
1
])(exp[iu
2
U
2
])(exp[iU
3
(u
1
Cu
2
))
Zexp[K(l
1
Cl
2
Cl
3
)Cl
1
exp[iu
1
]Cl
2
exp[iu
2
]
Cl
3
exp[i(u
1
Cu
2
)] (4)
We accept this as the denition of a bivariate Poisson
random variable. It can also be shown that [12]
P[N
1
ZjhN
2
Zl]
Z
1
4p
2
_
p
Kp
_
p
Kp
F
12
(u
1
; u
2
)exp[Ki(u
1
jCu
2
l)]du
1
du
2
Zexp[K(l
1
Cl
2
Cl
3
)]

min(j;l)
iZ0
l
jKi
1
l
lKi
2
l
i
3
(jKi)!(lKi)!i!
(5)
This result can be generalized for kO2 in a reasonably
simple manner. The number of mutually independent
Poisson random variables can be generalized to be given
by C
1
k
CC
2
k
where C
n
k
denotes combination of k random
variables taken n at a time. Thus, for kZ3, we consider
six mutually independent Poisson random variables U
i
with parameters l
i
(iZ1,2,.,6) and dene the trans-
formation
N
1
(a
1
; 0; T) =U
1
-U
4
-U
5
N
2
(a
2
; 0; T) =U
2
-U
4
-U
6
N
3
(a
3
; 0; T) =U
3
-U
5
-U
6
(6)
The equation relating l
i
(iZ1,2,.,6) with moments of
N
i
(a
i
,0,T), iZ1,2,3 can be shown to be given by
1 0 0 1 1 0
0 1 0 1 0 1
0 0 1 0 1 1
0 0 0 1 0 0
0 0 0 0 1 0
0 0 0 0 0 1
_

_
_

_
l
1
l
2
l
3
l
4
l
5
l
6
_

_
_

_
Z
(N
1
(a
1
;0;T))
(N
2
(a
2
;0;T))
(N
3
(a
3
;0;T))
Cov[N
1
(a
1
;0;T)N
2
(a
2
;0; T)]
Cov[N
1
(a
1
;0;T)N
3
(a
3
;0; T)]
Cov[N
2
(a
2
;0;T)N
3
(a
3
;0; T)]
_

_
_

_
(7)
For the case of kZ3, the joint characteristic function of
N
1
, N
2
and N
3
can be shown to be given by
F
123
(u
1
;u
2
;u
3
)
Zexp K

6
jZ1
l
j
C

3
jZ1
l
j
expiu
j
Cl
4
expi(u
1
Cu
2
Cu
3
)
_
Cl
5
expi(u
1
Cu
3
)Cl
6
expi(u
2
Cu
3
)
_
(8)
It should be noted here that in constructing this model
we need information on only the mean and covariance of
N
i
(a
i
,0,T), iZ1,2,.,k%n even when kO2. It is also to
be emphasized in this context that the assumption that
the number of crossings of a given level in a given
duration is Poisson distributed is not always valid.
Cramer [5] has shown that this assumption is asympto-
tically exact when the threshold levels increase to
innity. However, as pointed out by Vanmarcke [18,
19], for barrier levels of practical interest, this assump-
tion results in errors whose size and effect depend upon
the bandwidth of the underlying parent process. For wide
band processes, the Poisson approximation makes no
allowance for the time the process spends above the
threshold levels, while, for narrow band processes, the
level crossings tend to occur in clumps thereby
introducing statistical dependence between occurrence
of level crossings. Thus the models for extremes and
other order statistics developed in the present study are
not free of limitations resulting from Poisson approxi-
mations for the number of level crossings.
2.2. Passage times
For the purpose of illustration we begin by considering
the case of kZ1. It is well known that the rst passage time
T
11
(a) is related to the counting process N
1
(a,0,T) through
the relation
P[T
11
(a)OT] ZP[N
1
(a; 0; T) Z0] Zexp[Kl
1
T] (9)
Here it is assumed that P[T
11
(a)Z0]Z0. The above result
can be generalized to characterize the time for the nth
C.S. Manohar, R. Ghanem / Probabilistic Engineering Mechanics 20 (2005) 8796 89
crossing of level a with positive slope by X
1
(t). Thus one
gets
P[T
1n
(a)OT] ZP[N
1
(a; 0; T)%n K1]
Z

nK1
rZ0
exp(Kl
1
T)
(l
1
T)
r
r!
(10)
Based on this, we can also derive the joint distribution
P[T
1n
(a)OThT
1m
(a)OT]
ZP[N
1
(a; 0; T)%(nK1)hN
1
(a; 0; T)%(mK1)]
ZP[N
1
(a; 0; T)%min(nK1; mK1)] (11)
The extension of the above results for case of kO1 can be
achieved in a reasonably straightforward manner. Thus for
kZ2, one gets
P[T
ij
(a)OThT
rs
(b)OT]
ZP[N
i
(a; 0; T)!(j K1)hN
r
(b; 0; T)%(s K1)] (12)
This probability can be evaluated knowing the joint
distribution of the two counting processes (Eq. (5)).
2.3. Ordered peaks
Fig. 1 shows a sample of a Gaussian stationary random
process in which the positive peaks are marked with a star. It
is well known that Gumbel random variables serve as
acceptable models for the highest peak in a given time
interval. The parameters of this model can be derived in
terms of number of times a given level is crossed with
positive slope. This counting process can approximately be
modeled as a Poisson process [16]. This result can be
generalized on two fronts:
(a) Determination of the joint PDF of highest peaks of a
vector of Gaussian random processes.
(b) Determination of the joint PDF of ordered peaks, other
than just the highest, for a vector of Gaussian random
processes.
The rst of these generalizations has been discussed in
the recent study by Gupta and Manohar [10] in the
context of time variant system reliability analysis. In the
present study we focus attention on the second general-
ization. We again begin by considering the scalar case of
kZ1. The jth highest peak is related to the jth crossing
statistic by
P[X
1j
%a] ZP[T
1j
(a)OT] (13)
Thus, for the rst three highest peaks, one gets
P[X
11
%a] ZP[T
11
(a)OT] ZP[N
1
(a; 0; T) Z0]
Zexp(Kl
1
T)
P[X
12
%a] ZP[T
12
(a)OT] ZP[N
1
(a; 0; T)%1]
Zexp(Kl
1
T)[1 Cl
1
T]
P[X
13
%a] ZP[T
13
(a)OT] ZP[N
1
(a; 0; T)%2]
Zexp(Kl
1
T)[1 Cl
1
T C0:5(l
1
T)
2
]
(14)
The associated probability density functions (PDFs) can
be easily by derived by differentiation with respect to a. It
is of interest to note that in this formulation, ordered
peaks are mutually dependent. Thus, for instance
P[X
1r
%ahX
1(rCs)
%a]
ZP[N
1
(a; 0; T)%(r K1)hN
1
(a; 0; T)%(r Cs K1)]
ZP[N
1
(a; 0; T)%r K1] sP[X
1r
%a]P[X
1(rCs)
%a]
(15)
It is important to note that the proposed model imposes a
specic type of dependence structure on the ordered
peaks. To illustrate the formulation for the case of kO1,
we consider kZ2. Here one gets
P[X
1j
%ahX
2r
%b] ZP[T
1j
(a)OThT
2r
(b)OT]
ZP[N
1
(a; 0; T)%(j K1)hN
2
(b; 0; T)%(r K1)] (16)
This probability can be deduced knowing the joint PDF of
the counting processes N
1
and N
2
(Eq. (5)). Similar
extensions for kO2 are straightforward albeit requiring
involved formulary.
3. Numerical illustrations
3.1. Case 1
Here we consider a zero mean, scalar, stationary,
Gaussian random process X(t) with autocovariance and Fig. 1. Sample of X(t) with positive peaks shown.
C.S. Manohar, R. Ghanem / Probabilistic Engineering Mechanics 20 (2005) 8796 90
PSD functions given, respectively, by
R
xx
(t) Z2 exp(Ka
0
t
2
)cos(b
0
t); KN!t!N S
xx
(u)
Z

p
a
0
_
exp K
(uKb
0
)
2
4a
0
_ _
Cexp K
(uCb
0
)
2
4a
0
_ _ _ _
;
KN!u!N
(17)
It can be shown that [16]
(N
X
(a; 0; T)) Z
s
2
2ps
1
exp K
a
2
2s
2
1
_ _
s
2
1
Z2; s
2
2
Z2(b
2
C2a)
(18)
Fig. 2 shows the PDF of nth passage time for crossing a
threshold of 2s
1
for nZ1,2,3, and 4. The PDF of the rst
four ordered peaks over a time period of 0TZ42.4 s is
shown in Fig. 3 along with the rst order PDF of the parent
process. The time duration of 42.5 s here correspond to
about 25 times the correlation length of the process X(t). The
results on PDF of the rst four ordered peaks are compared
with corresponding results from Monte Carlo simulations
with 700 samples in Fig. 4. It can be observed from this
gure that there exists reasonable mutual agreement
between theoretical and simulation results with the agree-
ment being closer at the upper tails of the distribution
functions. This feature is consistent with the fact that the
assumption of Poisson model for N(a,0,T) is increasingly
valid for higher values of a. In these results it is assumed
that a
0
Z0.5, and b
0
Z2.0.
3.2. Case 2
Here we examine the case of two mutually dependent
and jointly stationary Gaussian random processes X(t)
and Y(t). We take X(t) to be identical with the scalar
random process considered in the previous case and dene
Y(t)ZA
1
X(t)CA
2
X(tC3). Fig. 5 shows the autocovariance,
Fig. 3. Probability density function of parent process and nth order statistics
of X(t); TZ42.4 s.
Fig. 4. Probability distribution of the rst four highest peaks; o denotes
results from a 700 samples Monte Carlo simulations; TZ42.4 s.
Fig. 2. Probability density function of nth passage times of X(t); aZ2s
1
.
Fig. 5. Normalized autocovariance and cross covariance of X(t) and Y(t).
C.S. Manohar, R. Ghanem / Probabilistic Engineering Mechanics 20 (2005) 8796 91
normalized with respect to variance, and cross correlation
coefcient of X(t) and Y(t) for A
1
Z0.3, A
2
Z0.7 and
3Z2.12 s. It can be observed that X(t) and Y(t) are away
from the extreme cases of being fully correlated or
completely uncorrelated. Figs. 6 and 7, respectively, show
the joint PDF of the rst highest peaks and the second
highest peaks of X(t) and Y(t) computed using Eq. (16).
Here time duration over which ordered peaks are studied
is TZ42.4 s which is about 25 times the correlation length
of the process X(t) and about 20 times the correlation
length of the process Y(t). For the four random variables
consisting of the rst two highest peaks of X(t) and Y(t),
respectively, the results on moments up to second order
were obtained using theory and Monte Carlo simulations
with 700 samples. The results on mean are found to be
3.56 (3.57), 2.98 (3.03), 2.68 (2.67), and 2.27 (2.30). Here
the numbers in the parenthesis correspond to the
simulation results. Similarly, the standard deviations of
the four random variables were obtained as 0.67 (0.64),
0.50 (0.46), 0.46 (0.49), and 0.35 (0.37). The correlation
coefcient matrix for these four random variables was
determined to be
r Z
1 0:63(0:61) 0:62(0:64) 0:45(0:47)
1 0:51(0:48) 0:59(0:56)
1 0:73(0:68)
Sym 1
_

_
_

_
The reasonable agreement that is found to exist between
the results of theory and simulations lends credence to the
formulations developed in this study.
3.3. Case 3
Here we consider the ordered peaks of response of linear
multi degree of freedom (MDOF) systems subject to random
excitations. Specically we focus on examining the
sensitivity of properties of ordered peaks of the response
processes with respect to changes in values of structural
parameters. This objective, in turn, is motivated by the
question if properties of ordered peaks could serve as
effective tools for detecting damage in the structure using
vibration signatures. We believe that such a possibility has
not been explored in the existing literature. The term
damage here is taken to denote any changes in structural
stiffness, mass or damping characteristics. We consider an
MDOF system in its undamaged state governed by the
equation
M

Z CC
_
Z CKZ ZF(t) (19)
The system is assumed to start from rest. The excitation
vector F(t) is taken to consist of zero mean, jointly
stationary Gaussian random processes. Upon the occurrence
of damage, the structure equation of motion is modied to
read
[MCDM]

Z CD

Z C[CCDC]
_
Z CD
_
Z
C[K CDK]Z CDZ ZF(t) (20)
Here DM, DC and DK denote the changes in mass, damping
and stiffness matrices due to occurrence of damage and DZ
is the concomitant change in the response. The matrix of
receptance functions for the structure, in its damaged state,
can be written as
H(u) CDH(u)
Z[Ku
2
(MCDM) Ciu(CCDC) C(K CDK)]
K1
(21)
Using the notation D(u)Z[Ku
2
MCiuCCK] and DD
(u)Z[Ku
2
DMCiuDCCDK]; we can write
H(u) CDH(u) Z[D(u) CDD(u)]
K1
(22)
Fig. 6. Joint probability distribution of highest peaks of X(t) and Y(t);
TZ42.4 s.
Fig. 7. Joint probability distribution of the second highest peaks of X(t) and
Y(t); TZ42.4 s.
C.S. Manohar, R. Ghanem / Probabilistic Engineering Mechanics 20 (2005) 8796 92
Using Neumanns expansion, we get
H(u) CDH(u) Z[I KP(u) C0:5P
2
(u) C.]D
K1
(u)
(23)
Here P(u)ZD
K1
(u)DD(u). Based on this, a second order
expansion for the PSD matrix of response of the damaged
structure can be obtained as
S(u) CDS(u) ZS(u) K[S(u)P
t+
(u) CP(u)S(u)]
C0:5[S(u)(P
t+
(u))
2
CP
2
(u)S(u)]
CP(u)S(u)P
t+
(u) (24)
Here the superscripts t and * denote, respectively, the matrix
transposition and conjugation. Furthermore, S(u) denotes
the PSD matrix of the response of the structure in its
undamaged state and is given by
S(u) ZD(u)S
FF
(u)D
t+
(u) (25)
Based on Eqs. (24) and (25) one can compute the properties
of ordered peaks of any desired response quantity for the
damaged and undamaged structures using formulary devel-
oped in Section 2. We limit our attention in this study to
investigate the relative magnitudes of changes in response
quantities due to a given set of damage scenarios. To
illustrate the formulation we consider the MDOF system
shown in Fig. 8. We take, for the undamaged structure,
M
1
Z1.5M, M
2
ZM, M
3
Z3M, M
4
Z2.5M, M
5
Z2M, K
1
ZK,
K
2
Z2K, K
3
Z1.5K, K
4
Z2K, K
5
Z3K, K
6
Z1.5K, K
7
Z2K
and K
8
Z4K with MZ10 kg and KZ1000 N/m. The natural
frequencies of this structure are determined to be 5.75,
11.77, 15.75, 22.19, and 28.50 rad/s. The damping in the
structure is taken to be viscous and proportional with
damping ratios h
1
and h
2
in the rst two modes being 0.03.
Furthermore, mass 3 is taken to be driven by a band-limited
white noise with a strength of unity and frequency range of
050 rad/s. To denote the changes in structural parameters
we introduce the notation
c ZM
1
; M
2
; .; M
5
; K
1
; K
2
; .; K
8
; h
1
; h
2

t
The structural parameters, after the occurrence of damage,
are denoted by c
i
CDc
i
Zc
i
x
i
(iZ1,2,.,15). We rst
consider the scenario in which
x Z0:99 0:98 0:97 0:97 0:98 0:97 0:98 0:99
0:99 0:97 0:99 0:98 0:98 0:97 0:99
This means that the structural damage is spatially
distributed and affects mass, stiffness and damping proper-
ties. In the discussion to follow we designate this damage
scenario as case 3.1. On account of these changes, the
structure natural frequencies change to 5.77, 11.84, 15.83,
22.13 and 28.48 rad/s. Fig. 9 shows the PSD of response at
the third mass for the undamaged and the damaged
structure. As can be observed the difference is hardly
perceptible. Also shown in this graph is the difference
between the two PSD functions. In these plots, the results on
Fig. 8. Randomly excited MDOF system.
Fig. 9. PSD of response for undamaged and damaged structures. The
difference of the two PSD functions is also shown.
C.S. Manohar, R. Ghanem / Probabilistic Engineering Mechanics 20 (2005) 8796 93
modied structure, obtained using the formulation in
Eq. (24), is compared with the simulation results. It can
be observed that the sensitivity analysis predicts the changes
in response fairly accurately. Figs. 10 and 11 show the
changes that are observed in the probability distribution of
passage times and the PDF of the ordered peaks. In
computing the passage times, a threshold of two times the
response standard deviation is set. Similarly, in computing
the order statistics a time duration of 20 s is used. Fig. 11
also shows the changes observed in the PDF of the parent
response process due to occurrence of damage. It is
interesting to note that the changes in the distribution of
higher order passage times are more pronounced than
similar results for lower order passage times. Conversely,
the changes produced in the density functions of the lower
order peaks are higher than that observed for the higher
order peaks. In any case, the changes in PDF of ordered
peaks are far higher than that observed in the PDF of the
parent process. This is, perhaps, to be expected, since, PDF
of higher order passage times and ordered peaks depends
upon many details of time evolution of response process
which a rst order PDF of the parent response process does
not explicitly take into account. It may also be noted that
Figs. 11 and 12 contain results from sensitivity analysis as
well as simulated results and these results are in reasonably
good mutual agreement. To illustrate the effect of isolated
changes to structural properties we consider the case when
one of the springs (K
1
, see Fig. 8) is changed by an amount
of 1% (case 3.2). Figs. 12 and 13, respectively, show the
results on differences on probability distribution of passage
times and the ordered peaks. These gures show trends that
Fig. 11. Difference in probability density function of nth order statistics of
response of undamaged and damaged structures; case 3.1; extended lines
denote theoretical results and o denotes simulation results; * denotes the
theoretical results for the parent process; the full line close to * denotes
simulation results; TZ20 s.
Fig. 12. Difference inPDFof nthpassage time for response at mass 3; case 3.2;
extended lines denote theoretical results and o denotes simulation results;
aZ2s.
Fig. 10. Difference in PDF of the nth passage time for response at mass 3;
case 3.1; extended lines denote theoretical results and o denotes
simulation results; sZ2s.
Fig. 13. Difference in probability density function of nth order statistics of
response of undamaged and damaged structures; case 3.2; extended lines
denote theoretical results and o denotes simulation results; * denotes the
theoretical results for the parent process; the full line close to * denotes
simulation results; TZ20 s.
C.S. Manohar, R. Ghanem / Probabilistic Engineering Mechanics 20 (2005) 8796 94
are similar to those observed in Figs. 10 and 11. To examine
the relative positioning of measurement point and the
location of damage, the choice of response DOF has been
varied and the resulting differences in distribution of
passage times and the density functions of the ordered
peaks are shown in Figs. 14 and 15. As one might expect,
responses at locations closer to the location of the damage
show relatively higher sensitivity with respect to the
damage.
4. Conclusions
The focus of the present study has been on characterizing
the probability distribution of ordered peaks of stationary
Gaussian random processes. Specically, the study has
explored the relationship between processes that count
the number of times a specied threshold level is crossed,
the times for crossing of a specied threshold level for the
nth time (nZ1,2,.) and the probability distribution of
ordered peaks. Using the results from multi-variate Poisson
processes, models for multi-variate probability distributions
of ordered peaks of vector Gaussian random processes have
been derived. The analytical predictions are shown to
compare reasonably well with limited results from Monte
Carlo simulations. One of the key features of the present
study has been the assumption of independent crossings that
is implicit in the Poissonian models for counting the level
crossings. Improvements to the model can be achieved by
using rened Markov models for the crossings. Similarly,
the extension of the procedures developed to handle
nonstationary and/or nonGaussian random processes require
further work. The present study also explores the sensitivity
of ordered peak characteristics of response processes of
randomly driven linear MDOF systems to small changes in
structural properties. Based on the illustrative example
considered, it is found that higher order passage times and
ordered peaks are more sensitive to structural changes than
the response process itself. This one might expect, since, the
examination of higher order passage times and ordered
peaks permits a scrutiny of response processes in greater
detail than the study of the probability structure of the
response process itself. We believe that this feature has
important applications in problems of structural model
updating and vibration signature based structural damage
detection. Such applications can be developed either using
inverse sensitivity procedures or by employing Bayesian
updation procedures. The present authors are currently
pursuing studies on these lines.
References
[1] Amini A, Trifunac MD. Distribution of peaks in linear earthquake
response. J Eng Mech ASCE 1981;107(EM1):20727.
[2] Basu B, Gupta VK, Kundu D. Ordered peak statistics through digital
simulation. Earthquake Eng Struct Dyn 1996;25(10):106173.
[3] Basu B, Gupta VK, Kundu D. A Markovian approach to ordered peak
statistics. Earthquake Eng Struct Dyn 1996;25(11):133551.
[4] Castillo E. Extreme value theory in engineering. Boston: Academic
Press; 1988.
[5] Cramer H. On the intersections between the trajectories of a normal
stationary stochastic process and a high level. Arkiv Math 1966;6:
33749.
[6] Galambos J. The asymptotic theory of extreme order statistics. New
York: Wiley; 1978.
[7] Gumbel EJ. Statistics of extremes. New York: Columbia University
Press; 1958.
[8] Gupta ID, Trifunac MD. A note on the statistics of ordered peaks in
stationary stochastic processes. Soil Dyn Earthquake Eng 1998;17:
31728.
[9] Gupta ID, Trifunac MD. Order statistics of peaks in earthquake
response. J Eng Mech ASCE 1988;114(10):160527.
[10] Gupta S, Manohar CS. Multivariate extreme value distributions for
random vibration applications, Under review. J Eng Mech ASCE;
2004.
Fig. 14. Difference in PDF of third passage time for response at masses 15;
case 3.2; aZ2s.
Fig. 15. Difference in probability density function of third order statistics of
response at masses 15 of undamaged and damaged structures; case 3.2;
TZ20 s.
C.S. Manohar, R. Ghanem / Probabilistic Engineering Mechanics 20 (2005) 8796 95
[11] Husler, J. Extreme values of nonstationary random sequences. J Appl
Probability 1986;23:93750.
[12] Johnson NL, Kotz S. Discrete distributions. New York: Wiley; 1969.
[13] Kotz S, Nadarajah S. Extreme value distributions. London: Imperial
College Press; 2000.
[14] Madsen HO, Krenk S, Lind NC. Methods of structural safety. New
Jersey: Prentice Hall; 1986.
[15] Nigam NC, Narayanan S. Applications of random vibrations. Berlin:
Springer; 1994.
[16] Nigam NC. Introduction to random vibrations. Boston: MIT Press;
1983.
[17] Resnick SI. Extreme values, regular variation, and point processes.
Berlin: Springer; 1987.
[18] Vanmarcke EH. Properties of spectral moments with applications to
random vibrations. J Eng Mech ASCE 1972;98:42546.
[19] Vanmarcke EH. On the distribution of the rst passage time for
normal stationary random processes. J Appl Mech ASME 1975;
March:21520.
C.S. Manohar, R. Ghanem / Probabilistic Engineering Mechanics 20 (2005) 8796 96