RECALL: Parameter Estimation: Lecture 15 Robust Estimation ...
Parameter Estimation for Mechanical Systems via an...
Transcript of Parameter Estimation for Mechanical Systems via an...
1
Parameter Estimation for Mechanical Systems via
an Explicit Representation of Uncertainty
Emmanuel Blanchard
Department of Mechanical Engineering, Virginia Polytechnic Institute and State
University, Blacksburg, Virginia, USA
Adrian Sandu
Computer Science Department, Virginia Polytechnic Institute and State University,
Blacksburg, Virginia, USA
Corina Sandu
Department of Mechanical Engineering, Virginia Polytechnic Institute and State
University, Blacksburg, Virginia, USA
Abstract
Purpose – To propose a new computational approach for parameter estimation in the Bayesian framework. Aposteriori
PDFs are obtained using the polynomial chaos theory for propagating uncertainties through system dynamics. The new
method has the advantage of being able to deal with large parametric uncertainties, non-Gaussian probability densities,
and nonlinear dynamics.
Design/methodology/approach - The maximum likelihood estimates are obtained by minimizing a cost function
derived from the Bayesian theorem. Direct stochastic collocation is used as a less computationally expensive alternative
to the traditional Galerkin approach to propagate the uncertainties through the system in the polynomial chaos
framework.
Findings – The new approach is explained and is applied to very simple mechanical systems in order to illustrate how the Bayesian cost function can be affected by the noise level in the measurements, by undersampling, non-identifiablily
of the system, non-observability, and by excitation signals that are not rich enough. When the system is non-identifiable
and an apriori knowledge of the parameter uncertainties is available, regularization techniques can still yield most likely
values among the possible combinations of uncertain parameters resulting in the same time responses than the ones
observed.
Originality/value – The polynomial chaos method has been shown to be considerably more efficient than Monte Carlo
in the simulation of systems with a small number of uncertain parameters. To the best of our knowledge, it is the first
time the polynomial chaos theory has been applied to Bayesian estimation.
Keywords - Parameter Estimation, Polynomial Chaos, Collocation, Halton/Hammersley Algorithm, Bayesian
Estimation, Vehicle Dynamics
Paper type Research paper
2
Nomenclature
A = matrix of basis function values at the collocation
points
#A = Moore-Penrose pseudo-inverse of A
ic = stochastic coefficient for the ith polynomial chaos
mode ( ix for displacement and iv for velocities)
0F = uncertain amplitude of the forcing function (N)
h = observation operator
H = matrix converting the states of the model to the
observable parameters of the system
kH = linearization of h about the solution ky
K = stiffness of the spring (N m-1
)
0K = mean value for the stiffness distribution (N m-1)
M = model solution operator which integrates the
model equations forward in time
M = mass of the body (kg)
0M = mean value for the mass distribution (kg)
n = number of random variables
On = number of observations
Pn = number of parameters
Sn = dimension of the state vector
N = number of time points at which measurements
are available
N = Normal distribution
p = maximum order of the polynomial basis
P = Probability Density Function (PDF)
d1P = 1-dimensional polynomial
Q = number of collocation points
kR = covariance matrix of the uncertainty associated
with the measurements noise at time step k
S = number of terms used in polynomial; chaos
expressions
t = time (s)
0t = initial time (s)
kt = time at the kth measurement point (s)
Ft = final time (s)
v = velocity (m s-1)
x = displacement (m)
ky = the state of the model at time kt
kz = observation vector at time kt
Z = second order random process
Greek symbols
k = measurement noise at time index k
i = ith collocation point
= set of uncertain parameters
= random event
)( = probability distribution of the multi-
dimensional random variable .
K = standard deviation for the stiffness distribution
M = standard deviation for the mass distribution
meas = standard deviation of the measurement
oscillation frequency obs
= frequency of the forcing function (radians s-1)
n = natural frequency (radians s-1)
obs = measured oscillation frequency (radians s-1)
= multi-dimensional random variable
= space of possible value for the unknown variables
= generalized Askey-Wiener polynomial chaoses
Subscripts
nom = nominal value, i.e., value obtained when 0
Superscripts
ref = reference value used to generate observations
3
1. Introduction and background
The polynomial chaos theory has been shown to be consistently more efficient than Monte Carlo
simulations in order to assess uncertainties in mechanical systems (Sandu et al., 2006a–2006b).
This paper extends the polynomial chaos theory to the problem of parameter estimation, and
illustrates is on a simple mass-spring system with uncertain initial velocity. We also present the
application of regularization techniques applied on this system with uncertain stiffness and
uncertain mass.
Parameter estimation is an important problem, because in many applications parameters cannot
be directly measured with sufficient accuracy; this is the case, for example, in real time
applications. Rather, parameter values must be inferred from available measurements of different
aspects of the system response. The theoretical foundations of parameter estimation can be found in
(Tarantola, 2004), (Bishwal, 2008) and (Aster et al., 2004). Parameter estimation find applications
in many fields, including mechanical engineering (Liu, 2008), material science (Araújo et al.,
2009), aerospace (Pradlwarter et al. 2005), geosciences (Catania and Paladino, 2009), chemical
engineering (Varziri et al., 2008), etc.
Various approaches to parameter estimations are discussed in the literature. These include
energy methods (Liang, 2007), frequency domain methods (Oliveto et al., 2008), and set inversion
via interval analysis (SIVIA) with Taylor expansions (Raїssi et al., 2004). A rigorous framework for
parameter estimation is the Bayesian approach, where probability densities functions are being
considered representations of uncertainty. The Bayesian approach has been widely used (Mockus et
al., 1997; Thompson and Vladimirov, 2005; Wang and Zabaras, 2005; Khan and Ramuhalli, 2008).
The Bayesian approach consists of estimating aposteriori probabilities of the parameters and
therefore transforming a parameter estimation problem into the problem of finding maximum
likelihood values of the parameters.
Different methodologies to estimate parameters in a Bayesian framework are possible.
Maximum likelihood parameter estimation can be formulated as an optimization problem (typically
large and nonconvex, therefore challenging). It can be numerically solved by gradient methods
(Nocedal and Wright, 2006) or by global optimization methods (Horst et al., 2000; Floudas, 2000;
Horst and Pardalos, 1995; Pardalos and Romeijn, 2002; Liberti and Maculan, 2006). Another
approach to solving the global continuous optimization problem is the use of Evolutionary
Algorithms (EAs) which are inspired by biological evolution (Davis, 1991). Differential Evolution
(DE) techniques are EA techniques that have been used successfully and Estimation of Distribution
Algorithms (EDAs) are a promising new class of EAs (Zhang, 1999). Sun et al. (2005) proposed a
DE/EDA hybrid approach. Another hybrid called estimation of distribution algorithm with local
search (EDA/L) has been developed by Zhang et al. (2004). Zhang et al. (2005) also proposed an
evolutionary algorithm with guided mutation (EA/G).
Another Bayesian parameter estimation method is the Kalman Filter (Kalman, 1960), which
is optimal for linear systems with Gaussian noise. The Extended Kalman Filter (EKF) allows for
nonlinear models and observations by assuming that the error propagation is linear (Evensen, G.,
1992, Evensen, G., 1993). The Ensemble Kalman Filter (EnKF) is a Monte Carlo approximation of
the Kalman filter suitable for large problems (Evensen, G., 1994). In the context of stochastic
optimization, propagation of uncertainties can be studied represented using Probability Density
Functions (PDFs). The Kalman Filter and its approximations estimate the states and their
4
uncertainties through covariance matrices at the same time. In order to approximate PDFs
propagated through the system, linearization using the EKF (Blanchard et al., 2007b) and Monte
Carlo techniques using the EnKF (Saad et al., 2007) are common approaches.
Parameter estimation is especially difficult for large systems, where a considerable
computational effort is needed. Estimating a large number of parameters often proves to be
computationally intractable. This has led to the development of techniques to determine which
parameters affect the system’s dynamics the most and are important to estimate (Sohns et al., 2006).
Sohns et al. (2006) proposed the use of activity analysis as an alternative to sensitivity-based and
principal component-based techniques. Their approach combines the advantages of the sensitivity-
based techniques (i.e., being efficient for large models) and the component-based techniques (i.e.,
keeping parameters that can be physically interpreted). Zhang and Lu (2004) combined the
Karhunen–Loeve decomposition and perturbation methods with polynomial expansions in order to
evaluate higher-order moments for saturated flow in randomly heterogeneous porous media.
In this study, we develop a maximum likelihood parameter estimation approach in a
Bayesian framework. Aposteriori PDFs are obtained using the polynomial chaos theory for
propagating uncertainties through system dynamics. The method has the advantage of being able to
deal with large parametric uncertainties, non-Gaussian probability densities, and nonlinear system
dynamics.
The polynomial chaos method started to gain attraction after Ghanem and Spanos (1990, 1991,
1993, 2003) applied it successfully to the study of uncertainties in structural mechanics and
vibration using Wiener-Hermite polynomials. Xiu extended the approach to general formulations
based on Wiener-Askey polynomials family (Xiu and Karniadakis, 2002a), and applied it to fluid
mechanics (Xiu et al., 2002; Xiu and Karniadakis, 2002b, 2003). Sandu et al. applied for the first
time the polynomial chaos method to multibody dynamic systems (Sandu et al. 2004, 2005, 2006a,
2006b), terramechanics (Li et al., 2005; Sandu et al., 2006c), and parameter estimation (Blanchard
et al., 2007a - 2007b). Saad et al. (2007) coupled the polynomial chaos theory with the Ensemble
Kalman Filter (EnKF) to indentify unknown variables in a non-parametric stochastic representation
of the non-linearities in a shear building model. Their identification method proved to be an
effective way of accurately detecting changes in the behavior of a system affected by both
measurement noise and modeling noise.
The generalized polynomial chaos theory is explained by Sandu et al. (2006 a) in which direct
stochastic collocation is proposed as a less expensive alternative to the traditional Galerkin
approach. It is desirable to have more collocation points than polynomial coefficients to solve for.
In that case a least-squares algorithm is used to solve the system with more equations than
unknowns.
The fundamental idea of polynomial chaos approach is that random processes of interest can be
approximated by sums of orthogonal polynomial chaoses of random independent variables. In this
context, any uncertain parameter can be viewed as a second order random process (processes with
finite variance; from a physical point of view they have finite energy). Thus, a second order random
process )(Z , viewed as a function of the random event , can be expanded in terms of orthogonal
polynomial chaos (Ghanem and Spanos, 2003) as:
1
))(()(j
jjcZ (1)
5
Here )( 1 n
j are generalized Askey-Wiener polynomial chaoses, in terms of the multi-
dimensional random variable n
n )( 1 , where is is the space of possible value for the
unknown variables. The Askey-Wiener polynomial chaoses form a basis with respect to the joint
probability density )( 1 n in the ensemble inner product
d)()()(, .
The basis functions are selected depending on the type of random variable functions. For
Gaussian random variables the basis functions are Hermite polynomials, for uniformly distributed
random variables the basis functions are Legendre polynomials, for beta distributed random
variables the basis functions are Jacobi polynomials, and for gamma distributed random variables
the basis functions are Laguerre polynomials (Xiu and Karniadakis, 2002a, 2003). In practice, a
truncated expansion of equation (1) is used,
S
j
jjcZ1
)( (2)
For independent random variables n 1 , the multi-dimensional basis functions are tensor products
of 1-dimensional polynomial basis:
plSj k
n
kk
l
kn
j k ,,2,1;,,2,1,)(d)1(P)(1
1
(3)
where !!
)!(
pn
pnS
, d)1(P is a 1-dimensional polynomial, n is the number of random variables, and p is
the maximum order of the polynomial basis. The total number of terms S increases rapidly with n
and p .
In the deterministic case, a second order system can be described by the following Ordinary
Differential Equation (ODE):
00
00
)(
)(
),(),(,)(
)(
vtv
xtx
tvtxtFtv
vtx
(4)
In the stochastic framework developed in this study, the displacement )(tx , the velocity )(tv , and the
set of parameters Pn (the set of Pn parameters being possibly uncertain) of a second order
system can be expanded using equation (2) as:
pS
S
i
ii
ll
S
i
ii
mm
S
i
ii
mm nlnmtvtvtxtx
1,1,)()(,)()(),(,)()(),(111
(5)
We use subscripts to index system components and superscripts to index stochastic modes. Inserting
Eq. (5) in the deterministic system of equations leads to:
0,0
1111
0
)(
)(;)(,)(,)()(
,1,1),()(
mm
S
k
kkS
k
kkS
k
kk
m
jS
j
j
m
FS
j
m
j
m
xtx
vxtFtv
tttSjnmtvtx
(6)
6
where m is a state index and Sn is the dimension of the state vector.
To derive evolution equations for the stochastic coefficients )(tx i
m we impose that equation (6)
holds at a given set of collocation points Q ,,1 . This leads to:
QiAvAxAtFvAvx kS
k
ki
kS
k
ki
kS
k
ki
j
m
S
j
ji
i
m
i
m
1,;,,,1
,
1
,
1
,
1
,
(7)
where A represents the matrix of basis function values at the collocation points:
SjQiAAA ij
jiji 1,1),(, ,, (8)
The collocation points have to be chosen such that SQ and
A is nonsingular. Let #A be the
Moore-Penrose pseudo-inverse of
A . With kS
kki
i xAX
1
,, k
S
kki
i vAV
1
,, k
S
kki
i A
1
,, the collocation
system can be written as:
QiVXtFVVX iiiiii 1,,,,, (9)
After integration, the stochastic solution coefficients are recovered using:
SitVAtvtXAtxS
j
j
ji
ijS
jji
i
1,)()(),()(1
,
#
1,
# (10)
Assuming that 1 is the constant (zeroth order) term in the polynomial expansion, the mean values
of )(tx and )(tv are 11 )()( txtx and 11 )()( tvtv , respectively. The standard deviations of )(tx and
)(tv are given by:
dtxS
i
ii
2
2
)()(
, dtv
S
i
ii
2
2
)()(
(11)
When the basis functions are orthogonal polynomials, the standard deviations of )(tx and )(tv are
given by:
iiS
i
i tx ,)(2
2
, iiS
i
i tv ,)(2
2
(12)
When the basis are orthonormal, the standard deviations of )(tx and )(tv are given by:
S
i
i tx2
2)( ,
S
i
i tv2
2)( (13)
The Probability Density Functions (PDF) of )(tx and )(tv are obtained by drawing histograms of
their values using a Monte Carlo simulation and normalizing the area under the curves that are
obtained. It is not computationally expensive since the Monte Carlo simulation is run on the final
result, and not for the whole process. For instance, the ODE is run the same number of times as the
number of collocation points, which is typically much lower than the number of runs used for the
Monte Carlo simulation.
7
2. Bayesian approach for parameter estimation
Optimal parameter estimation combines information from three different sources: the physical laws
of evolution (encapsulated in the model), the reality (as captured by the observations), and the
current best estimate of the parameters. The information from each source is imperfect and has
associated errors. Consider the mechanical system model (6) which advances the state in time
represented in a simpler notation:
Nktyyytyv
x
y kkk
k
k
k
k ,,2,1,,,, 0011
M
(14)
The state of the model Sn
ky at time moment kt depends implicitly on the set of parameters
Pn , possibly uncertain (the model has Sn states and Pn parameters). M is the model solution
operator which integrates the model equations forward in time (starting from state 1ky at time 1kt
to state ky at time kt ). N is the number of time points at which measurements are available.
For parameter estimation it is convenient to formally extend the model state to include the
model parameters and extend the model with trivial equations for parameters (such that parameters
do not change during the model evolution)
1 kk (15)
The optimal estimation of the uncertain parameters is thus reduced to the problem of optimal state
estimation. We assume that observations of quantities that depend on system state are available at
discrete times kt
k
T
kkkkkkkkk RyHyhz ,0, (16)
where On
kz is the observation vector at kt , h is the (model equivalent) observation operator, kH
is the linearization of h about the solution ky , and k is the measurement noise at time index k .
Note that there are On observations for the Sn –dimensional state vector, and that typically SO nn .
Each observation is corrupted by observational (measurement and representativeness) errors (Cohn,
1997). We denote by the ensemble average over the uncertainty space. The observational error
is the experimental uncertainty associated with the measurements and is usually considered to have
a Gaussian distribution with zero mean and a known covariance matrix kR .
Using polynomial chaoses the uncertain parameters are modeled explicitly as functions of a set
of random variables p with a joint probability density function . The explicit
dependency of the system state on the random variables is obtained via a collocation approach:
S
i
ii
kk
S
i
ii tyty11
)(,, (17)
We adopt the point of view that the “state of knowledge” about the uncertain parameters can be
described by probability densities. From Bayes’ rule the probability density of the parameter
distribution conditioned by all observations is
8
dyzzyPyzP
zzyPyzPzzyP
NN
NNNN
NN][][
][][
01
01
0
(18)
where ][ NN yzP is the Probability Density Function (PDF) of the latest observational error (taken
at time Nt ), ]|[ 01 zzyP NN is the “model forecast PDF” conditioned by all previous observations
(taken at times 0t to 1Nt ), and ]|[ 0zzyP NN is the “assimilated PDF”. The assimilated PDF
represents the aposteriori probability of the parameters after all the observations have been taken
into account.
For simplicity denote by y the current state of the system (the best estimation obtained using all
previous observations 01 zzN ) and by Nzz the latest, yet-to-be-used set of observations.
Moreover, consider that the observational error has a Gaussian distribution with covariance kR and
that the observations at different times are independent. Then Bayes’ formula becomes
n
N
k
kkkT
kk
N
k
kkkT
kk
n
dyyPe
yPe
dyyPyzP
yPyzPzyP
yHzRyHz
yHzRyHz
][
][
][][
][][
0
1
21
0
1
21
)()(
)()(
(19)
The unconditional probability density ][yP is the PDF of the current system state, and is
implicitly represented by the polynomial chaos expansion of the state yy . Moreover,
integration against this probability density can be evaluated by integration in the independent
random variables
fdyfdyyPyfn anyfor)(][)( (20)
The denominator can be evaluated by a multidimensional integration. However, in our
approach, there is no need to evaluate this scaling factor, since its omission does not change the
estimation procedure. (The omission of this scaling factor is equivalent to adding a constant to the
function we minimize, and this does not change the result of the minimization procedure).
The mean of the best state estimate that uses the new observations z is obtained from Bayes
formula as
deydyyPeydyzyPyy
N
k
kkkT
kk
n
N
k
kkkT
kk
n
yHzRyHzyHzRyHz
)()(den
1][
den
1]|[ 0
1
2
1
0
1
2
1)()()()(
(21)
For the parameter estimation the Bayes’ formula specializes to:
m
N
k
kkkT
kk
N
k
kkkT
kk
dPe
PePzPzP
yHzRyHz
yHzRyHz
][
][
den
][][
0
1
21
0
1
21
(22)
Note that the aposteriori probability defined by Bayes formula can be written (in principle) as a
function of the independent random variables
9
de
ezP
N
k
kkkT
kk
N
k
kkkT
kk
yHzRyHz
yHzRyHz
)(
ˆ
0
1
21
0
1
21
(23)
In this setting polynomial chaos is used to model the a priori pdf of the parameters; the Bayes
formula is employed to obtain the a posteriori pdf (i.e., the pdf conditioned by the observations).
The maximum likelihood estimate is given by that realization of the parameters (that value of
) which maximizes the a posteriori probability ]|[ zP , or, equivalently, minimizes ]|[log zP :
)(log)()(min0
1
21
N
kkkk
T
kk yHzRyHzJ (24)
Note that for we have 0 and cost the function J becomes infinite. This cost function
is composed of two parts:
apriori
mismatch
)(log)()(0
1
21
total
JJ
N
kkkk
T
kk yHzRyHzJ
(25)
where mismatchJ comes only from the differences between the available measurements and the model
response, while aprioriJ encapsulates the apriori knowledge of the parameter uncertainty. The value
Jminargˆ minimizing the cost function (25) gives the most likely values of our uncertain
parameters as ˆˆ .
3. Insight into the Bayesian approach using simple mechanical systems
We now illustrate the proposed Bayesian approach for the estimation of parameters of several
simple mechanical systems. We discuss how the cost function and the estimate can be affected by
low sampling rates (i.e., below the Nyquist frequency), by measurement noise, and by non-
identifiability issues.
The state of the model n
ky at time moment kt depends implicitly on the set of uncertain
parameters p , and therefore on the set of independent random variables . This dependency is
explicitly represented in the polynomial chaos framework, specifically, at each time moment kt the
state is given as a polynomial of the random variables kk yy . The probability density of the state
can also be obtained from this relation. H is simply a matrix converting the states of the model
)(y to the observable parameters of the system (i.e., the quantities which can be measured), which
are contained in z . kR is the covariance matrix of the uncertainty associated with the
measurements, i.e., of the measurement noise.
)(log aprioriJ comes from the apriori knowledge of the uncertain parameters. Using
polynomial chaoses the uncertain parameters can be modeled explicitly as functions of a set of
random variables p with a joint probability density function .
10
mismatchJ is usually the most important component of the cost function, but aprioriJ is useful when
mismatchJ does not contain enough information in order to find a clear minimum value for our cost
function. This is illustrated in the next section of this article.
3.1 Mass-spring system with uncertain initial velocity
This section applies the Bayesian approach to the simple mass-spring system shown in Figure 1.
K
M
x
K
M
x
Figure 1. Mass–spring system
The parameters K (the stiffness of the spring) and M (the mass of the body) are known. The
system has a zero initial displacement 00 x but a nonzero initial velocity 0v (e.g., created by
hitting the mass from below with a hammer at 0t , which will produce 00 v ). We want to
estimate the uncertain initial condition 0v based on measurements of the displacement )(tx at later
times.
The equation of motion of the system is
0)()( M txKtx (26)
and admits the general solution (Inman, 2001):
0
01
2
2
0
2
0 tansin)(v
xt
xvtx n
n
n
n
,
M
Kn (27)
The values chosen for the numerical experiments are N/m9478.3 )(20.1 K 2 , kg 0.1 M , and
therefore Hz 1rad/s π2 n . For these values, and with 00 x , the solution (27) becomes:
.2cos)(,2sin2
)( 00 tvtvt
vtx
(28)
The amplitude of )(tx and the amplitude of )(tv are both proportional to the uncertain parameter 0v ,
as illustrated in Figure 2. A single measurement of the displacement at a time 21 mt (with m
integer) allows to estimate the initial velocity as )2sin()(2 110 ttxv . Note that 0)2sin( 1 t . A
single measurement of both the velocity and the displacement at any time 2t is sufficient to retrieve
the initial velocity, since for any 2t at least one of the variables is nonzero, 0)2sin( 2 t or
0)2cos( 2 t .
11
Figure 2. Displacements and velocities of the mass–spring system
We now consider the case where measurements of the displacement only are taken at multiple
time moments Nttt ,,, 21 . This will give insight on how the two parts of the Bayesian cost function
can be affected by low sampling rates (i.e., below the Nyquist frequency) and by measurement
noise.
We assume some prior knowledge of the initial velocity, which represents how hard different
people can hit the mass with the hammer. The range of possible initial velocities is between m/s 5.0
and m/s 5.1 , with a most likely value of m/s 10 v . We model this prior knowledge as shown in
Figure 3. Let be a random variable with a Beta(2,2) probability distribution )( in the range
]1,1[ . The random initial velocity is then
.]/[5.01)0,( 0 smvv nom (29)
Figure 3. Beta (2, 2) distribution for 0v
12
The state of the system at future times depends on the random initial velocity and can be
represented by
),(
),(
),(
),(
t
tv
tx
ty
. Synthetic measurements are obtained from a reference simulation
with the reference value of the uncertain parameter 23.0ref . If we assume that only the
displacement can be measured we have that 001H and the measurements yield
.,0,)()( refref
kkkkkkk RtxtyHz Ν (30)
The measurement noise k is assumed to be Gaussian with a zero mean and a variance 1% (or
0.01% or 10% when indicated) of the value of )(tx plus the maximum measured value of )(tx
divided by 1000. Therefore, the covariance of the uncertainty associated with the measurements is
][)(max001.001.0,10max 2212 mzzR ktkk . This value is always greater than zero and 1
kR can
always be computed. Measurement errors at different times are independent random variables.
The maximum likelihood estimate is obtained by minimizing the Bayesian cost function
apriori
mismatch
)(log),(),(1
1
21
total
JJ
N
kkkk
T
kk tHyzRtHyzJ
(31)
The random system output ),( ty is discretized using 6 terms in the polynomial chaos expansions,
and 12 collocation points will be used to derive the polynomial chaos coefficients. The collocation
points used in this study are obtained using an algorithm based on the Halton algorithm (Halton and
Smith, 1964), which is similar to the Hammersley algorithm (Hammersley, 1960).
The frequency of the output signal ),( tx is 1 Hz for any value of . If )(tx is measured every
s5.0 from s5.0t to s5t , then 0),( ktx for any value of and kkz mismatch part gives no
extra information, as shown in Figure 4, in which the plot for )(tx was obtained with 23.0 .
N
kkk
T
k RJ1
1
2
1mismatch (32)
For clarity we will illustrate this detailing the step by step procedure using analytical formulas for
this particular example.
The explicit dependency of ),( tx is obtained via a collocation approach. It can be represented as
6
1
)()(),(S
i
ii txtx (33)
The equation of motion of the system illustrated in Figure 1 is given by equation (26). As shown
earlier, the solution for this equation for Hz 1rad/s π2 M
Kn and for a zero initial
displacement is given y equation (28).
In a polynomial chaos framework, the equation of motion of the system yields six equations for
6S :
6,,2,1,0)()( SitxKtxM ii (34)
13
Solving the six different equations of motions separately yields
6,,2,1,2sin2
)( 0 Sitv
tx k
i
i
(35)
and the polynomial expression of the displacement is
6
1
0 )(2sin2
),(S
i
i
k
i
tv
tx
(36)
Let’s note that, in the general case, there is no need to know the closed form solution of the
equations of motion. The approach presented in this paper still works when using the states
variables obtained with numerical techniques to solve ODE’s at the chosen collocation points.
The coefficients iv0 are obtained using
66
0
55
0
44
0
33
0
22
0
11
000 5.0)0,( vvvvvvvvv nomnom (37)
For Beta(2,2) distributed random variables the basis are Jacobi (1,1) polynomials. With one random
variable and for the range ]1,1[ , the normalized Jacobi (1,1) polynomials are:
) 132 330- 300 120- 20(-1 )(7/4455028
) 42 84- 56 14-(1 ) (3/92822
) 14 21- 9(-1 (5/8344)
) 5 5-(1 1/106)(
) 2(-1 3/28)(
1/2
54326
4325
324
23
2
1
(38)
Therefore, the coefficients iv0 are
0,0,0,0,3
7,
2
56
0
5
0
4
0
3
0
02
0
01
0 vvvvv
vv
vnomnom
(39)
The cost function from equation (31) can be written as
)(log)(2sin2
1
2
1)(
1
26
1
0
N
k
i
k
S
i
i
k
k
tv
zR
J , 214
3)( (40)
Using the fact that only 1
0v and 2
0v are nonzero and replacing 1
0v by nomv0)2/5( and 2
0v by nomv0)3/7( ,
it yields
2
1
2
001
4
3log) 2(-1 3/28)(2sin
2
)3/7(
2
12sin
2
)2/5(1
2
1
N
kk
nom
k
nomk
k
tv
tv
zR
J (41)
which can be simplified as
2
1
2
001
4
3log) (22sin
2
)4/(2sin
2
1
2
1
N
kk
nom
k
nomk
k
tv
tv
zR
J (42)
which is also equal to
14
22
1
2
0
1
00
1
2
0
14
3log22sin
2
)4/(1
2
1
22sin2
)4/(2sin
2
12sin
2
1
2
1
N
kk
nom
k
N
kk
nom
k
nomk
k
N
kk
nomk
k
tv
R
tv
tv
zR
tv
zR
J
(43)
The closed form solution of the value minimizing the total cost function is a long expression that is
not written here. The mismatch part of the cost function is
2
1
2
0
1
2
0
2
0
1
2
0
22sin2
)4/(1
2
1
22sin2
2sin2
1
2sin2
1
2
1
N
kk
k
N
kkk
k
k
N
kk
k
k
mismatch
tv
R
tv
tv
zR
tv
zR
J
(44)
The value mismatchmismatch Jminargˆ minimizing mismatch part the cost function
N
kk
k
N
kkk
k
k
mismatch
tv
R
tv
tv
zR
1
2
0
1
2
0
2
0
2sin2
)4/(12
2sin2
2sin2
1
ˆ
(45)
If )(tx is measured every 0.5 s from 5.0t to 5.1t , then 0),( ktx for any value of and kkz
N
k k
k
mismatchR
J1
2
2
1 (46)
which is the formula that was already obtained in equation (32)
In this case, the denominator of equation (45) is not defined and mismatch is not defined, because the
mismatch part does not depend on and yields an estimation where all possible values of
5.0100 vv are equally likely. Therefore, the value totalJminargˆ minimizing the total cost
function is also the value minimizing the apriori part of the cost function, i.e., 0ˆ .
3.2 Possible impact of undersampling on the cost function
Increasing the number of measurements generally yields a better estimation, and as a general rule,
sampling above the Nyquist frequency rate should always be done when possible. This section
studies the possible impact of undersampling for two reasons. Measurements might not be available
at a rate above the Nyquist frequency rate when this frequency is very high, for instance. It will also
be shown that in some cases, it is possible to know that the estimation is already quite accurate
when using only a very few measurements points instead of all of them. This can be useful when
computational time is an issue and an answer is needed quickly, which does not prevent one from
continuing to process the extra information later on if needed, knowing that the extra measurements
will generally yield more precision.
15
The mismatch part of the cost function is driven by the observational errors. The summed
contribution of errors makes mismatchJ a random variable with a 2 distribution with n degrees of
freedom. For a relatively large number of measurements this distribution behaves like a normal one
with mean n and variance n. The mismatch part does not depend on and yields an estimation
where all possible values of 0v are equally likely. This is illustrated in Figure 4 where the mismatch
part of the cost function is constant. The mismatch part does not depend on the noise level in this
case since 1
kR is inversely proportional to k
T
k . In Figure 4, as well as in consequent figures,
we identify with the “ ο ” sign the points where measurements were available (i.e., the points for
which we collected measurements).
In this case the estimation relies entirely on the apriori part of the cost function, i.e. )(log .
The estimate coincides with the best initial guess, i.e., 0ˆ . This is really the worst-case scenario:
the frequency of sampling the output is below the Nyquist frequency rate, and the sampling points
are exactly those time moments when the displacement is zero and the observations contain no
information.
(a) (b) Estimate = 0
Figure 4. Bayesian estimation with 10 time points
Notes: a) Displacement when no noise added, b) Estimation with noise = 1%
If we measure )(tx at any other additional time point then the Bayesian approach yields an
accurate estimation result (for any reasonable amount of measurement noise). We can interpret this
fact as follows. If the output sampling is not done at least at the Nyquist frequency, one cannot
guarantee that all the relevant information in the output signal is captured, i.e. one cannot guarantee
that the mismatch part of the cost function will bring extra information. We still have a PDF of the
possible values of the uncertain parameter, but in the worst case scenario, it will be no better than
the apriori PDF.
In most practical situations, however, it is very likely that the Bayesian approach will provide
an accurate estimate even when the output is sampled below the Nyquist frequency. In the example
above a single measurement point is sufficient, provided that the measurement time is not one for
which the displacement is zero. This is where the Parameter Estimation and Signal Reconstruction
16
differ. In the above setting of parameter estimation one samples outputs of the system. If the outputs
were arbitrary signals then their full reconstruction would require a sufficient sampling frequency.
But the outputs are constrained by the input and by the system dynamics, and only a small set of all
the possible reconstructed signals are consistent with both the known input and with the equations
of motion. The reconstruction of signals in this small family requires less information than the
reconstruction of arbitrary signals. In our example all possible system outputs form a one-parameter
family of signals (frequency of 1 Hz, phase equal to zero, and variable amplitude). Consequently a
single measurement of the output is almost always sufficient to estimate the single uncertain
parameter.
(a) (b) Estimate = 0.23
(c) Estimate = 0.23 (d) Estimate = 0.19
Figure 5. Bayesian estimation with 3 time points
Notes: a) Displacement when no noise added, b) Estimation with noise = 0.01%,
c) Estimation with Noise = 1%, d) Estimation with noise = 10%
The practical question is now how to decide whether the sampling of the output is sufficient.
The answer is given by the shape of the Bayesian cost function which indicates whether there is
17
enough information to obtain a good estimate or not. The second derivative of the cost function at
the minimum approximates the inverse of the covariance of the uncertainty in the estimate. Loosely
speaking, the sharper the minimum of the cost function the more trustworthy the estimate is; and the
wider the minimum the larger the estimation error can be.
The role of the shape of the cost function is illustrated in Figure 5, in which only three
measurements points for 0t are used. Different levels of measurement errors lead to different
shapes of the cost function, and to different estimation accuracies. For noise levels of 0.01% and
0.1% the total cost function is almost equal to its mismatch part for all values of , it has a sharp
minimum, and the Bayesian approach yields an accurate estimate. For very noisy measurements
(10%) the relative weight of the information coming from measurements is smaller, and the relative
weight of the apriori information is higher. Consequently the apriori part of the cost function is
more significant and the minimum of the total cost function is wider. In this scenario the output
sampling is done below the Nyquist frequency, but we know that the estimates are accurate for
noise levels of 0.01% and 1% because the cost functions have clear minima.
(a) (b) Estimate = 0.23
(c) Estimate = 0.23 (d) Estimate = 0.22
Figure 6. Bayesian estimation with 30 time points
18
Notes: a) Displacement when no noise added, b) Estimation with noise = 0.01%,
c) Estimation with noise = 1%, d) Estimation with noise = 10%
One very accurate output sample would be enough for a perfect estimation. Taking more sample
points leads to a better estimation for noisy measurements because the effect of the noise averages
out as we take more samples. Figure 6 illustrates the cost function when 30 measurements are used.
The relative weight of the mismatch part increases and we get a better estimation when the noise is
10%.
When the extra samples do not bring additional information to the estimation process the cost
function changes as shown in Figure 7. The net result of the additional measurements is to add a
constant to the mismatch part (which corresponds to the effect of measurement noise). The shape of
the cost function does not change, in particular the minimum is not more pronounced, and the
quality of the estimate is not improved.
(a) (b) Estimate = 0
(c) (d) Estimate = 0
Figure 7. Effect of adding sample points containing no useful information
19
Notes: a) Displacement when no noise added with 5 time points, b) Estimation with 5 time points and noise = 0.01%,
c) Displacement when no noise added with 10 time points, d) Estimation with 10 time points and noise = 0.01%
Next we consider the situation where both )(tx and )(tv are measured at times 0kt . The
observation operator is
010
001H and the measured values are
v
k
x
k
k
k
kkktv
txtyHz
),(
),(),(
. The measurement noise is assumed Gaussian with zero mean and covariance matrix
212
212
))((max001.0)(01.0,10max0
0))((max001.0)(01.0,10max
tvtv
txtxR
tk
tkk
The inverse 1
kR can always be computed. One data point at any 0t is sufficient to estimate our
unknown parameter 0v for low noise levels, as shown in Figure 8. In the general case, however,
measurements of the full state vector do not guarantee that they contain useful information when the
sampling rate is below the Nyquist frequency.
(a) (b)
(c) Estimate = 0.23 (d) Estimate = 0.19
Figure 8. Bayesian estimation with 1 time point when velocity measurements are available
20
Notes: a) Displacement when no noise added, b) Velocity when no noise added,
c) Estimation with noise = 0.01%, d) Estimation with noise = 1%
3.3 Mass-spring system with sinusoidal forcing function
This section applies the parameter estimation Bayesian approach developed in this study to the
simple mass-spring system with sinusoidal forcing function shown in Figure 9.
K
M
x
t)(ω cos FF(t) 0
K
M
x
t)(ω cos FF(t) 0
Figure 9. Mass –Spring System with sinusoidal forcing function
The parameters K (the stiffness of the spring), M (the value of the mass) are known. The
system is initially at equilibrium, i.e., it has zero initial displacement 00 x and velocity 00 v . The
problem is to estimate the uncertain amplitude of the forcing function 0F .
We assume the following apriori information. 0F has a Beta(2,2) distribution in the range
]1500,500[ NN with the most likely value N 000,10 F . With ]1,1[ a Beta(2,2) distributed
variable the apriori distribution of 0F is
NNF 500000,10 (47)
The reference value of the force amplitude is NF 115,1ref
0 , or 23.0ref . This reference value is
used to generate artificial observations and is not available to the estimation procedure. The
numerical values of the other parameters are as follows: N/m765,17 )2(1.5200 2 K , kg 200 M ,
00 x , 00 v , and Hz 0.5rad/s π . Note that Hz 1.5rad/s π21.5 MKn .
The equation of motion of the system is:
)(cos)()( M 0 tFtxKtx (48)
The solution is sought in the time interval from t = 0 to t =5. Since 00 x and 00 v , the analytical
solution of this equation of motion is (Inman, 2001):
tt
MFtx nn
n2
sin2
sin)/(2
)(22
0
(49)
With our numerical values, the displacement of the mass can be written as:
ttFtx 2sinsin101.2665)( 0
-4 (50)
21
It can be seen that the amplitude of )(tx and the amplitude of )(tv are both proportional to the
uncertain parameter 0F , as shown in Figure 10. Therefore, the estimation of 0F can be in principle
based on a single measurement of the output.
Figure 10. Displacement and velocities of the Mass – Spring system with sinusoidal forcing function
(a) (b)
22
(c) Estimate = 0 (d) Estimate = 0.23
Figure 11. Bayesian estimation with 5 time points when velocity measurements are available
Notes: a) Displacement when no noise added, b) Velocity when no noise added,
c) Estimation with noise = 0.01%, d) Estimation with noise = 10%
Figure 11 illustrates the effect of measurements of both )(tx and )(tv at five time points. This
sampling provides no information on the uncertain parameter, and in this worst-case scenario the
estimate is based solely on apriori information. A sampling of the output below the Nyquist
frequency does not guarantee that we get sufficient information from the output signal about the
uncertain parameter. The mismatch part does not depend on the noise level in this case since 1
kR is
inversely proportional to k
T
k , as shown in equation (32).
(a) (b)
(c) Estimate = 0.23 (d) Estimate = 0.25
Figure 12. Bayesian estimation with 3 time points when velocity measurements are available
a) Displacement when no noise added, b) Velocity when no noise added,
c) Estimation with noise = 0.01%, d) Estimation with noise = 10%
23
However, we can see in Figure 12 that 3 time measurements yield an accurate estimation for a
low noise level, even though the sampling is well below the Nyquist frequency. Once again, the
shape of the cost function indicates that for low noise levels we have enough information to
accurately estimate our uncertain parameter. While there are many signals with a maximum
frequency of 1.5 Hz which fit the observations at the chosen three measurement times, only one of
them is consistent with the input signal and with equation of motion.
3.4 Regularization techniques applied to a mass-spring system with uncertain stiffness and
uncertain mass
This example addresses the issue of non-identifiability. The Bayesian approach is applied to the
simple mass-spring system shown in Figure 1. The difference between the example in this section
and the one from Section 3.1 is that the uncertain parameters are different: they are the stiffness of
the spring ( K ) and the value of the mass ( M ). Our apriori information about the uncertain
parameters is expressed in terms of probability densities as follows. The mass has a normal
distribution with mean kg 2000 M and a standard deviation kg 6.667M . The stiffness has also a
normal distribution with mean N/m7,895.68 N/m 22002
0 K and standard deviation
N/m8.052,1K . We represent the uncertain parameters as functions of a random vector of two
independent normal random variables 21, as follows
.1,0,,, 212010 KM KKMM (51)
We consider the “true” values of the parameters to be kg533.201ref M and N/m62.7495ref K ,
which correspond to the reference values of the random variables )38.0,23.0(, ref
2
ref
1
ref .
These values are not available to the estimation process, but are used in a reference simulation to
generate synthetic observations.
We measure the values of the oscillation frequency obs (along the reference solution) and use
it to derive information about K and M . The measurement errors are assumed to have a normal
distribution with zero mean (unbiased) and a standard deviation equal to
rad/s 3938.001.0 00 MKmeas .
This example allows an analytical solution to the Bayesian approach and provides insight into
the role of the mismatch and the a priori parts of the cost function in the estimation. The position of
the mass is given by:
0
01
2
2
0
2
0 tansin)(v
xt
xvtx n
n
n
n
, (52)
and clearly it depends only on refref MKn .
The Bayesian cost function is defined in this case as:
24
2
2
obs10202
2
2
1
2
2
obs
2
2
0
2
2
0
)()(
2
1
2
1
2
1
)()(
2
1))((
2
1))((
2
1
meas
MK
measKM
MK
MKKKMMJ
(53)
The maximum likelihood value of the parameters is the argument that minimizes this cost function,
Jminargˆ . The contour plots shown in Figure 13 represent the mismatch part of the cost
function, its apriori part, and the total value of the cost function in the space of random variables.
(a) (b)
(c)
Figure 13. Contours of the cost function
Notes: a) mismatch part, b) apriori part, c) total cost function
The magnitude of the apriori part is relatively small and the total cost function is roughly equal to
its mismatch part. As expected, the mismatch part yields a line of possible minima, because the
measured ratio MKn does not contain information about the individual values of K and M .
25
The point ref is plotted in Figure 13 and it lies on the line of minima. The Bayesian interpretation
is the following: all the pairs ),( MK along the line are equally likely to produce the value of the
measured oscillation frequency. We say that the (individual values of the) parameters K and M are
non-identifiable.
When multiple combinations of uncertain parameter values result in the same observed behavior
of the system (same measurements) a regularization approach (Aster et al., 2004) can be used in
estimation. In order to find the most likely parameter values one increases the relative importance of
the apriori knowledge of the system. This is done by multiplying the apriori part by a
“regularization coefficient” large enough so that the new total cost function has a clear minimum
value along the possible values.
2
2
obs
2
2
0
2
2
02dregularize)()(
2
1))((
2
1))((
2
1,
measKM
MKKKMMJ
(54)
The net effect of regularization in this example is to reduce the standard deviations in the apriori
distributions (to M and K respectively), therefore to increase the trust in the apriori
information. The contour plots of the regularized cost functions are shown in Figure 14 for different
regularization coefficients.
Figure 14. Contour plots of the cost function after regularization for different coefficients
The cost function looks like its mismatch part when the regularization coefficient is very low
and looks like its apriori part when the regularization coefficient is very high. As the regularization
coefficient gets larger, the line of possible minima becomes an ellipse, and starts moving away from
the location of the original line of minima and toward (0, 0), the apriori most likely value. When the
line becomes an ellipse with a well defined center, the regularization coefficient is large enough; it
should not be further increased as this leads to an increase of the bias in the estimate. In our
example 62 105 seems to be a good value for the regularization coefficient and it results in the
estimated values )40.0,09.0(),( 21 . The choice of the regularization coefficient is problem-
dependent and requires a careful analysis of the resulting estimates. Regularization leads to biased
estimates, as stronger assumptions are being artificially imposed.
3.5 Discussion of the Bayesian approach
26
The quality of the maximum likelihood estimate is related to the shape of the Bayesian cost
function, with a sharp minimum indicating an accurate estimate. Inaccurate estimates can be caused
by different factors, including a sampling rate below the Nyquist frequency, non-identifiability,
non-observability, and an excitation signal that is not rich enough.
The parameters are non-identifiable when different parameter values lead to identical system
outputs. In this case the Bayesian cost function has an entire region of minima (e.g., a valley), with
each parameter value in the region being equally likely. A regularization approach based on
increasing the weight of the apriori information can be used to select reasonable estimates. When
non-observability is the reason leading to non–identifiability, the problem can be addressed by
including measurements of additional states in the estimation procedure, if possible. When an input
signal that is not rich enough is the reason leading to non –identifiability, the problem can be
addressed by changing the kind of excitations applied to the system
For identifiable and observable systems accurate estimates can be obtained in most cases even
if the output signal is sampled below the Nyquist rate. In the worst case, however, sampling below
the Nyquist rate cannot guarantee that sufficient information is extracted from the output. In this
worst case the apriori information becomes important and the estimate is biased toward the apriori
most likely value.
4. Summary and Conclusions
This paper applies the polynomial chaos theory to the problem of parameter estimation, using direct
stochastic collocation. The maximum likelihood estimates are obtained by minimizing a cost
function derived from the Bayesian theorem. This approach is applied to very simple mechanical
systems in order to illustrate how the cost function can be affected by undersampling, non-
identifiablily of the system, non-observability, and by excitation signals that are not rich enough.
Inaccurate estimates can be caused by those different factors. It has been shown that the quality of
the maximum likelihood estimate is related to the shape of the Bayesian cost function, with a sharp
minimum indicating an accurate estimate. The parameters are non-identifiable when different
parameter values lead to identical system outputs. In this case the Bayesian cost function has an
entire region of minima (e.g., a valley), with each parameter value in the region being equally
likely. Regularization techniques can still yield most likely values among the possible combinations
of uncertain parameters resulting in the same time responses than the ones observed. This was
illustrated using a simple spring-mass system. For identifiable and observable systems accurate
estimates can be obtained in most cases even if the output signal is sampled below the Nyquist
frequency. In the worst case, however, sampling below the Nyquist rate cannot guarantee that
sufficient information is extracted from the output. In this worst case the apriori information
becomes important and the estimate is biased toward the apriori most likely value.
The proposed method has several advantages. Simulations using Polynomial Chaos methods are
much faster than Monte Carlo simulations. Another advantage of this method is that it is optimal; it
can treat non-Gaussian uncertainties since the Bayesian approach is not tailored to any specific
distribution. We plan to apply the proposed technique to identify parameters of a real mechanical
system for which laboratory measurements are available.
Acknowledgements
27
This research was supported in part by NASA Langley through the Virginia Institute for
Performance Engineering and Research award. The authors are grateful to Dr. Mehdi Ahmadian,
Dr. Steve Southward, Dr. John Ferris, Dr. Sean Kenny, Dr. Luis Crespo, Dr. Daniel Giesy, and Mr.
Carvel Holton for many fruitful discussions on this topic.
References
Araújo, A.L., Mota Soares, C.M., Herskovits, J., Pedersen, P. (2009), “Estimation of Piezoelastic and
Viscoelastic Properties in Laminated Structures”, Composite Structures, Vol. 87, No. 2, pp. 168-174.
Aster, R.C., Borchers, B., and Thurber, C.H. (2004), Parameter Estimation and Inverse Problems, Elsevier Academic Press. ISBN 0-12-065604-3.
Bishwal, J.P.N. (2008), Parameter Estimation in Stochastic Differential Equations, Springer, Berlin,
Germany.
Blanchard, E., Sandu, C., and Sandu, A. (2007a), “A Polynomial-Chaos-based Bayesian Approach for
Estimating Uncertain Parameters of Mechanical Systems”, Proceedings of the ASME 2007 International
Design Engineering Technical Conferences & Computers and Information in Engineering Conference
IDETC/CIE 2007, 9th International Conference on Advanced Vehicle and Tire Technologies (AVTT), Las Vegas, Nevada, September.
Blanchard, E., Sandu, A., and Sandu, C. (2007b), “Parameter Estimation Method Using an Extended Kalman
Filter”, Proceedings of the Joint North America, Asia-Pacific ISTVS Conference and Annual Meeting of Japanese Society for Terramechanics, Fairbanks, Alaska, June.
Catania, F., and Paladino, O. (2009), “Optimal Sampling for the Estimation of Dispersion Parameters in Soil
Columns Using an Iterative Genetic Algorithm”, Environmental Modelling & Software, Vol. 24, No. 1, pp. 115-123.
Cohn, S. E. (1997), “An Introduction to Estimation Theory”, Journal of the Meteorological Society of Japan,
Vol. 75 (B), pp. 257-288.
Davis, L. (1991), Handbook of Genetic Algorithms, Van Nostrand Reinhold, New York.
Evensen, G. (1992), “Using the Extended Kalman Filter with a Multi-layer Quasi-geostrophic Ocean
Model”, Journal of Geophysical Research, Vol. 97 No. C11, pp. 17905-17924.
Evensen, G. (1993), “Open Boundary Conditions for the Extended Kalman Filter with a Quasi-geostrophic Mode”, Journal of Geophysical Research, Vol. 98 No C19, pp.16529-16546.
Evensen, G. (1994), “Sequential Data Assimilation with a Non-linear Quasi-geostrophic Model using Monte
Carlo Methods to Forecast Error Statistics” Journal of Geophysical Research, Vol. 99 No C5, pp. 10143–10162.
Floudas, C.A. (2000), Deterministic Global Optimization: Theory, Methods, and Applications, Kluwer
Academic Publishers, Dordrecht, The Netherlands.
Ghanem, R.G., and Spanos, P.D. (2003), Stochastic Finite Elements, Dover Publications Inc, Mineola, NY.
Ghanem, R.G., and Spanos, P.D. (1990), “Polynomial Chaos in Stochastic Finite Element”, Journal of Applied Mechanics, Vol. 57, pp. 197-202.
Ghanem, R.G., and Spanos, P.D. (1991), “Spectral Stochastic Finite-Element Formulation for Reliability Analysis”, ASCE Journal of Engineering Mechanics, Vol. 117 No. 10, pp. 2351-2372.
28
Ghanem, R.G., and Spanos, P.D. (1993), “A Stochastic Galerkin Expansion for Nonlinear Random Vibration Analysis”, Probabilistic Engineering Mechanics, Vol. 8 No. 3, pp. 255-264.
Halton, J. H., Smith, G. B. (1964), “Radical-inverse Quasi-random Point Sequence”, Communications of the
ACM, Vol. 7 No. 12, pp. 701–702.
Hammersley, J. M. (1960) “Monte Carlo Methods for Solving Multivariables Problems”, Annals of the New
York Academy of Sciences, Vol. 86, pp. 844–874.
Horst, R., and Pardalos, P.M., editors (1995), Handbook of Global Optimization, Vol. 1. Kluwer Academic
Publishers, Dordrecht, The Netherlands.
Horst, R., Pardalos, P. M., and Thoai, N. V. (2000), Introduction to Global Optimization, Second Edition,
Kluwer Academic Publishers, Dordrecht, The Netherlands.
Inman, D. J. (2001), Engineering Vibration, Second Edition, Prentice Hall, Inc., ISBN 0-13-726142-X.
Kalman, R. E. (1960), “A New Approach to Linear Filtering and Prediction Problems”, Transaction of the
ASME- Journal of Basic Engineering, pp. 35-45.
Khan, T., and Ramuhalli, P. (2008), “A Recursive Bayesian Estimation Method for Solving Electromagnetic Nondestructive Evaluation Inverse Problems”, IEEE Transactions on Magnetics, Vol. 44 No. 7, pp. 1845-
1855.
Li, L., Sandu, C., and Sandu, A. (2005), “Modeling and Simulation of a Full Vehicle with Parametric and
External Uncertainties”, Proceedings of the 2005 ASME Int. Mechanical Engineering Congress and Exposition, 7th VDC Annual Symposium on "Advanced Vehicle Technologies", Session 4: Advances in
Vehicle Systems Modeling and Simulation, Paper number IMECE2005-82101, Orlando, FL, November.
Liberti, L., and Maculan, N. (2006), Global Optimization: From Theory to Implementation, Springer, Berlin, Germany.
Mockus, J., Eddy, W., Mockus, A., Mockus, L., and Reklaitis, G. (1997), Bayesian Heuristic Approach to
Discrete and Global Optimization: Algorithms, Vizualization, Software and Applications, Kluwer Academic Publishers, Dordrecht, The Netherlands.
Nocedal, J., and Wright, S.J. (2006), Numerical Optimization, Vol.2, Springer.
Oliveto, N.D., Scalia, G., and Oliveto, G. (2008), “Dynamic Identification of Structural Systems with
Viscous and Friction Damping”, Journal of Sound and Vibration, Vol. 318, pp. 911–926.
Pardalos, P.M., and Romeijn, H.E. (2002), editors, Handbook of Global Optimization, Vol.2., Kluwer
Academic Publishers Dordrecht, The Netherlands.
Pradlwarter, H.J., Pellissetti, M.F., Schenk, C.A., Schuëller, G.I., Kreis, A., Fransen, S., Calvi, A., and Klein, M. (2005), “Realistic and Efficient Reliability Estimation for Aerospace Structures”, Computer Methods in
Applied Mechanics and Engineering, Special Issue on Computational Methods in Stochastic Mechanics and
Reliability Analysis, Vol. 194 No. 12-16, pp. 1597-1617.
Raїssi, T., Ramdani, N., and Candau, Y. (2004), “Set Membership State and Parameter Estimation for Systems Described by Nonlinear Differential Equations”, Automatica, Vol. 40, pp. 1771–1777.
Saad, G., Ghanem, R.G., and Masri, S. (2007), “Robust system identification of strongly non-linear
dynamics using a polynomial chaos based sequential data assimilation technique”, Collection of Technical Papers - 48
th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference,
Vol. 6, pp. 6005-6013, Honolulu, HI, May.
29
Sandu, A., Sandu, C., and Ahmadian, M. (2006a), “Modeling Multibody Dynamic Systems with
Uncertainties. Part I: Theoretical and Computational Aspects”, Multibody System Dynamics, Publisher: Springer Netherlands, ISSN: 1384-5640 (Paper) 1573-272X (Online), DOI 10.1007/s11044-006-9007-5, pp.
1-23 (23), June 29, 2006.
Sandu, C., Sandu, A., and Ahmadian, M. (2006b), “Modeling Multibody Dynamic Systems with
Uncertainties. Part II: Numerical Applications”, Multibody System Dynamics, Publisher: Springer Netherlands, ISSN: 1384-5640 (Paper) 1573-272X (Online), DOI: 10.1007/s11044-006-9008-4, Vol. 15, No.
3, pp. 241 - 262 (22), April 2006.
Sandu, C., Sandu, A., Chan, B.J., and Ahmadian, M. (2004), “Treating Uncertainties in Multibody Dynamic Systems using a Polynomial Chaos Spectral Decomposition”, Proceedings of the ASME IMECE 2004, 6th
Annual Symposium on “Advanced Vehicle Technology”, Paper number IMECE2004-60482, Anaheim, CA,
November.
Sandu, C., Sandu, A., Chan, B.J., and Ahmadian, M. (2005), “Treatment of Constrained Multibody Dynamic
Systems with Uncertainties”, Proceedings of the SAE Congress 2005, Paper number 2005-01-0936, Detroit,
MI, April.
Sandu, C., Sandu, A., and Li, L. (2006c), “Stochastic Modeling of Terrain Profiles and Soil Parameters”,
SAE 2005 Transactions Journal of Commercial Vehicles, Vol. 114 No. 2, pp. 211-220.
Sohns, B., Allison, J., Fathy, H.K., Stein, J.L. (2006), “Efficient Parameterization of Large-Scale Dynamic
Models Through the Use of Activity Analysis”, Proceedings of the ASME IMECE 2006, Chicago, Illinois, November.
Sun, J., Zhang, Q., Tsang, E.P.K. (2005) “DE/EDA: A New Evolutionary Algorithm for Global
Optimization”, Information Sciences, Vol. 169 No 3-4, pp. 249-262.
Tarantola, A. (2004), Inverse Problem Theory and Methods for Model Parameter Estimation, Society for
Industrial and Applied Mathematics.
Thompson, B., Vladimirov, I. (2005), “Bayesian Parameter Estimation and Prediction in Mean Reverting Stochastic Diffusion Models”, Nonlinear Analysis, Vol. 63, pp. e2367 – e2375.
Varziri, M.S., Poyton, A.A., McAuley, K.B., McLellan, P.J., and Ramsay, J.O. (2008), “Selecting Optimal
Weighting Factors in iPDA for Parameter Estimation in Continuous-Time Dynamic Models”, Computers &
Chemical Engineering, Vol. 32 No. 12, pp. 3011-3022.
Wang, J., Zabaras, N. (2005), “Using Bayesian Statistics in the Estimation of Heat Source in Radiation”,
International Journal of Heat and Mass Transfer, Vol. 48, pp. 15-29.
Xiu, D., Lucor, D., Su, C.H., and Karniadakis, G.E. (2002), “Stochastic Modeling of Flow-Structure Interactions using Generalized Polynomial Chaos”, Journal of Fluids Engineering, Vol. 124, pp. 51-59.
Xiu, D., and Karniadakis, G.E. (2002a), “The Wiener-Askey Polynomial Chaos for Stochastic Differential Equations”, Journal on Scientific Computing, Vol. 24 No. 2, pp. 619-644.
Xiu, D., and Karniadakis, G.E. (2003), “Modeling Uncertainty in Flow Simulations via Generalized Polynomial Chaos”, Journal of Computational Physics, Vol. 187, pp. 137-167.
Xiu, D., and Karniadakis, G.E. (2002b), “Modeling Uncertainty in Steady-state Diffusion Problems via Generalized Polynomial Chaos”, Computer Methods in Applied Mechanics and Engineering, Vol. 191, pp. 4927-4928.
Zhang, B.T. (1999), “A Bayesian Framework for Evolutionary Computation”, Proceedings of the 1999 Congress on Evolutionary Computation, Vol. 1, pp. 722-728.
30
Zhang, D., Lu, Z. (2004), “An Efficient, High-order Perturbation Approach for Flow in Random Porous
Media via Karhunen–Loeve and Polynomial Expansions”, Journal of Computational Physics, Vol. 194 No. 2, pp. 773–794.
Zhang, Q., Sun, J., Tsang, E., and Ford, J. (2004), “Hybrid Estimation of Distribution Algorithm for Global
Optimization”, Engineering Computations, Vol. 21 No. 1, pp. 91-107.
Zhang, Q., Sun, J., Tsang, E. P. K. (2005), “An Evolutionary Algorithm with Guided Mutation for the Maximum Clique Problem”, IEEE Transactions on Evolutionary Computation, Vol. 9 No. 2, pp. 192-200.
List of Figures
Fig. 1 Mass–spring system
Fig. 2 Displacements and velocities of the mass–spring system
Fig. 3 Beta (2, 2) distribution for 0v
Fig. 4 Bayesian estimation with 10 time points: a) Displacement when no noise added, b) Estimation with noise = 1%
Fig. 5 Bayesian estimation with 3 time points: a) Displacement when no noise added, b) Estimation with
noise = 0.01%, c) Estimation with noise = 1%, d) Estimation with noise = 10%
Fig. 6 Bayesian estimation with 30 time points: a) Displacement when no noise added, b) Estimation with
noise = 0.01%, c) Estimation with noise = 1%, d) Estimation with noise = 10%
Fig. 7 Effect of adding sample points containing no useful information: a) Displacement when no noise added with 5 time points, b) Estimation with 5 time points and noise = 0.01%, c) Displacement when
no noise added with 10 time points, d) Estimation with 10 time points and noise = 0.01%
Fig. 8 Bayesian Estimation with 1 time point when velocity measurements are available: a) Displacement when no noise added, b) Velocity when no noise added, c) Estimation with noise = 0.01%, d)
Estimation with noise = 1%
Fig. 9 Mass–Spring system with sinusoidal forcing function
Fig. 10 Displacement and velocities of the Mass–Spring system with sinusoidal forcing function
Fig. 11 Bayesian estimation with 5 time points when velocity measurements are available: a) Displacement
when no noise added, b) Velocity when no noise added, c) Estimation with noise = 0.01%, d) Estimation with noise = 10%
Fig. 12 Bayesian estimation with 3 time points when velocity measurement is available: a) Displacement
when no noise added, b) Velocity when no noise added; c) Estimation with noise = 0.01%, d) Estimation with noise = 10%
Fig. 13 Contours of the cost function: a) mismatch part, b) apriori part, c) total cost function