A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 1
PSM: A Polynomial Chaos-Based Kalman Filter Approach for
Parameter Estimation of Mechanical Systems
Emmanuel D. Blanchard*
Virginia Polytechnic Institute and State University, Department of Mechanical Engineering, 3103
Commerce Street, Blacksburg, VA 24061, USA
phone: (540) 231-0700, fax: (540) 231-0730, e-mail: [email protected]
Dr. Adrian Sandu
Virginia Polytechnic Institute and State University, Computer Science Department,
2224 Knowledge Works, Blacksburg, VA 24061, USA
phone: (540) 231-2193, fax: (540) 231-9218, e-mail: [email protected]
Dr. Corina Sandu
Virginia Polytechnic Institute and State University, Mechanical Engineering Department, 104
Randolph Hall, Blacksburg, VA 24061, USA
phone: (540) 231-7467, fax: (540) 231-9100, e-mail: [email protected]
ABSTRACT
Background. Mechanical systems operate under parametric and external excitation uncertainties. The polynomial chaos
approach has been shown to be more efficient than Monte Carlo for quantifying the effects of such uncertainties on the
system response. Many uncertain parameters cannot be measured accurately, especially in real time applications.
Information about them is obtained via parameter estimation techniques. Parameter estimation for large systems is a
difficult problem, and the solution approaches are computationally expensive.
Method of Approach. This paper proposes a new computational approach for parameter estimation based on the
Extended Kalman Filter (EKF) and the polynomial chaos theory for parameter estimation. The error covariances needed by EKF are computed from polynomial chaos expansions, and the EKF is used to update the polynomial chaos
representation of the uncertain states and the uncertain parameters. The proposed method is applied to a nonlinear four
degree of freedom roll plane model of a vehicle, in which an uncertain mass with an uncertain position is added on the
roll bar.
Results. The main advantages of this method are an accurate representation of uncertainties via polynomial chaoses, a
computationally efficient update formula based on EKF, and the ability to provide aposteriori probability densities of
the estimated parameters. The method is able to deal with non-Gaussian parametric uncertainties. The paper identifies
and theoretically explains a possible weakness of the EKF with approximate covariances: numerical errors due to the
truncation in the polynomial chaos expansions can accumulate quickly when measurements are taken at a fast sampling
rate. To prevent filter divergence we propose to lower the sampling rate, and to take a smoother approach where a set of
time-distributed observations are all processed at once.
Conclusions. We propose a parameter estimation approach that uses polynomial chaoses to propagate uncertainties and
estimate error covariances in the EKF framework. Parameter estimates are obtained in the form of a polynomial chaos
expansion which carries information about the aposteriori probability density function. The method is illustrated on a
roll plane vehicle model.
Keywords. Parameter Estimation, Polynomial Chaos, Collocation, Halton/Hammersley Algorithm, Extended Kalman
Filter (EKF), Vehicle Dynamics
*Corresponding Author
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 2
1. INTRODUCTION AND BACKGROUND
The polynomial chaos approach has been shown to be computationally more efficient than Monte Carlo for quantifying
uncertainties in mechanical systems [1, 2]. This paper extends the polynomial chaos theory to the problem of parameter
estimation, which is very relevant to physical system modeling since it first requires modeling uncertainties in order to
use the physical response to improve the model itself once the unknown parameters have been estimated. The proposed method is illustrated on a nonlinear four degree of freedom roll plane vehicle model, in which an uncertain mass and its
uncertain position are estimated.
Parameter estimation is an important problem, because in many instances parameters cannot be physically
measured, or cannot be measured with sufficient accuracy 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 [3-5]. Parameter estimation find applications in many fields, including mechanical
engineering [6], material science [7], aerospace [8], geosciences [9], chemical engineering [10], etc. A literature review
specific to the online estimation of onroad vehicles’ mass can be found in [11], in which an algorithm providing
conservative error estimates is also proposed
Various approaches to parameter estimations are discussed in the literature. These include energy methods [12],
frequency domain methods [13], and set inversion via interval analysis (SIVIA) with Taylor expansions [14]. 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 [15-18]. 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 [19] or by global optimization methods [20-24]. Another
approach to solving the global continuous optimization problem is the use of Evolutionary Algorithms (EAs) which are
inspired by biological evolution [25]. 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 [26]. Sun et al. [27]
proposed a DE/EDA hybrid approach. Another hybrid approach called estimation of distribution algorithm with local
search (EDA/L) has been developed by Zhang et al. [28]. Zhang et al. [29] also proposed an evolutionary algorithm with guided mutation (EA/G).
Another Bayesian parameter estimation method is the Kalman Filter [30], 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 [31, 32]. The Ensemble Kalman Filter (EnKF) is a Monte Carlo approximation of the Kalman
filter suitable for large problems [33]. In the context of stochastic optimization, propagation of uncertainties can be
represented using Probability Density Functions (PDFs). The Kalman Filter and its approximations estimate the states
and their uncertainties at the same time through covariance matrices. In order to approximate PDFs propagated through
the system, linearization using the EKF [34] and Monte Carlo techniques using the EnKF [35] are common approaches.
The EKF has the advantage of taking nonlinear dynamic effects into account and therefore dealing with non-Gaussian
probabilities, but the EnKF is more practical when dealing with large state space systems for which the covariance
matrix becomes too large. Particle filters are ensemble-based assimilation methods which can also take nonlinear
dynamic effects into account and deal with non-Gaussian probabilities, but are not adapted to high-dimensional systems [36].
Parameter estimation is well recognized as a theoretically difficult problem; moreover, estimating a large number
of parameters is often computationally very expensive. This has led to the development of techniques determining
which parameters affect the system’s dynamics the most, in order to choose the parameters that are important to
estimate [37]. Sohns et al. [37] 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.,
efficiency for large models) and of the component-based techniques (i.e., using parameters that can be physically
interpreted). Zhang and Lu [38] 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.
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 3
The polynomial chaos method started to gain attention after Ghanem and Spanos [39-42] 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 [43], and applied it to fluid mechanics
[44-46]. Sandu et al. applied for the first time the polynomial chaos method to multibody dynamic systems [1, 2, 47,
48], terramechanics [49, 50], and parameter estimation in the time domain for fixed parameters [51, 52]. In their
groundbreaking work, Soize and Ghanem [53] described mathematical settings for characterizing problems for which random uncertainties have arbitrary probability densities. Desceliers et al. [54] used a polynomial chaos representation
of a random field to be identified, developed a method to estimate the coefficients of that representation, and extended it
to apply it to experimental vibration tests using frequency response functions [55]. Saad et al. [56] 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. Li and Xiu [57] also developed a methodology combining the polynomial chaos theory with the EnKF,
in which they sampled the polynomial chaos expression of the stochastic solution in order to reduce the sampling errors.
The benefits and drawbacks of the EnKF are discussed in [58] and [59]. Smith et al. [60] designed a polynomial chaos
observer for indirect measurements which provides a full probability density function from the polynomial chaos
coefficients, and which is computationally less expensive than using a regular EKF. Their approach is designed to
compensate for the modeling noise but needs to be tuned (e.g. with a Kalman approach) to take observation noise into account. Finally, let us mention that long-time integration errors are a major problem with the polynomial chaos theory,
which has been addressed by Wan and Karniadakis [61] who developed a multi-element generalized polynomial chaos
(ME-gPC) method.
The generalized polynomial chaos theory developed by Xiu [43] is also explained by Sandu et al. [1] 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 relation between collocation and Galerkin methods is
explained in [1]. Cheng and Sandu [62, 63] further discuss the computational cost of using the polynomial chaos theory
with both Galerkin and collocation methods. The dimensionality of the problem decreases the efficiency of the
polynomial chaos theory as will be shown in Eq. (3).
The fundamental idea of the 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 )(X , viewed as a function of the random event , can be
expanded in terms of orthogonal polynomial chaos [39] as:
1
))(()(j
jjcX (1)
Here )( 1 n
j are generalized Wiener- Askey polynomial chaoses [64, 65], in terms of the multi-dimensional
random variable n
n )( 1 . The Wiener- Askey polynomial chaoses form a basis that is orthogonal with
respect to the joint probability density )( 1 n in the ensemble inner product
d, (2)
The multi-dimensional basis functions are tensor products of 1-dimensional polynomial bases:
bk
n
kk
l
kn
j plSjP k ,,2,1;,,2,1,)()(1
1
(3)
where !!
)!(
b
b
pn
pnS
, n is the number of random variables, and
bp is the maximum order of the polynomial basis.
The total number of terms S increases rapidly with n and bp .
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
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 4
gamma distributed random variables the basis functions are Laguerre polynomials [43, 46]. In practice, a truncated
expansion of Eq. (1) is used,
S
j
jjcX1
)( (4)
Consider a deterministic second order system which can be described by a system of ordinary differential
equations (ODE):
00
00
)(
)(
),(),(,)(
)()(
vtv
xtx
tvtxtFtv
tvtx
(5)
For unconstrained mechanical systems xntx )( represents the vector of displacements, xn
tv )( is the vector
of velocities, and Pn is the vector of parameters. In the stochastic framework developed in this study the uncertain
displacements, velocities, and parameters are expanded using Eq. (4) as:
px
S
i
ii
pp
S
i
ii
mm
S
i
ii
mm npnmtvtvtxtx
1,1,)()(,)()(),(,)()(),(111
(6)
Subscripts are used to index system components and superscripts are used to index stochastic modes. Inserting Eq. (6)
in the deterministic system of equations leads to:
Fmm
S
k
kkS
k
kkS
k
kk
m
jS
j
j
m
x
j
m
j
m
tttxtx
vxtFv
Sjnmvx
00,0
1111
,)(
)(;)(,)(,)(
1,1,
(7)
To derive evolution equations for the stochastic coefficients )(tx i
m , Eq. (7) is imposed to hold at a given set of
collocation points Q ,,1 . This leads to:
QiAvAxAtFvAvx kS
kki
kS
kki
kS
kki
j
m
S
jji
i
m
i
m
1,;,,,1
,1
,1
,1
, (8)
where A represents the matrix of basis function values at the collocation points:
SjQiAAA ij
jiji 1,1),(, ,, (9)
The collocation points have to be chosen such that SQ and
A has full rank. 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,,,,, (10)
After integration of these Q independent versions of the deterministic system, the stochastic solution coefficients are
recovered using:
.1,)()(),()(1
,
#
1,
# SitVAtvtXAtxQ
j
j
ji
ijQ
jji
i
(11)
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:
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 5
dtx iiS
i
i
)(),()(2
2, dtv ii
S
i
i )(),()(2
2
(12)
Similarly, the covariance of two variables can be computed from the polynomial chaos expansion. For example the
covariance of uncertainties in state component m and in parameter p is:
dtxtx iiS
i
i
p
i
mpm
)(),()(),(cov2
(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 obtained. In order to generate the PDF at any
time, random samples with an appropriate distribution need to be drawn and plugged into the polynomial chaos
representation of the time-dependent state. With the known coefficients and the random numbers, an ensemble of states can be generated and represented by a PDF. With the polynomial chaos approach, as long as the polynomial chaos
coefficients and bases are known, a large state ensemble can easily be generated to form a smooth PDF curve. However,
in a Monte Carlo approach, each member of the ensemble states requires a full system run. Therefore, generating a PDF
with the polynomial chaos approach is not computationally expensive, since the Monte Carlo simulation is run on the
final result, which corresponds to repeated evaluations of polynomial values but not repeated ODE simulations. To be
specific, the number of ODE runs equals the number of collocation points, which is typically much lower than the
number of runs used in a Monte Carlo simulation.
2. EXTENDED KALMAN FILTER 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 (7) which advances the state in time represented in a simpler notation:
Nktyyytyv
xy kkk
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 the system state are available at discrete times kt
kkkkkkkk RyHyhz ,0, N (16)
where On
kz is the observation vector at kt , h is the (model equivalent) observation operator and kH is the
linearization of h about the solution ky . 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
[66]. The observational error k 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 .
The Kalman filter [30-32, 67] assumes that the model (14) is linear, and the model state at previous time 1kt is
normally distributed with mean a
ky 1 and covariance matrix a
kP 1 . The Extended Kalman Filter (EKF) allows for
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 6
nonlinear models and observations by assuming that the error propagation is linear. The nonlinear observation operators
are also linearized.
The state is propagated from 1kt to kt using model equations (14), and the covariance matrix is explicitly propagated
using the tangent linear model operator 'M and its adjoint
M'*,
QPP a
k
f
k
*
1 '' MM (17)
where the superscripts f and a stand for “forecast” and “assimilated”, respectively. Q represents the covariance of the
model errors.
Under linear, Gaussian assumptions, the PDFs of the forecast and assimilated fields are also Gaussian, and completely
described by the mean state and the covariance matrix. The assimilated state a
ky and its covariance matrix a
kP are
computed from the model forecast f
ky , the current observations
zk , and from their covariances using:
.
,1
1
f
kk
T
k
f
kkk
T
k
f
k
f
k
a
k
f
kkk
T
k
f
kkk
T
k
f
k
f
k
a
k
PHHPHRHPPP
yHzHPHRHPyy
(18)
One step of the extended Kalman filter can be represented as:
kk
a
k
a
k
f
k
f
k
a
k
a
k
Rz
PyPyPy
and
andandandFilterModelLinearTangent&Model
11 (19)
For parameter estimation, the model state is extended to formally include the model parameters:
a
k
a
k
a
kk
f
k
f
k yty
1
111 ,,
M (20)
The covariance matrix of the extended state vector can be estimated from the polynomial chaos expansions of y and
.
f
k
f
k
f
k
f
k
f
k
f
k
f
k
f
kf
ky
yyyP
cov,cov
,covcovcov (21)
Using this covariance matrix, the Kalman gain matrix is computed using the formula:
1 T
k
f
kkk
T
k
f
kk HPHRHPK (22)
The Kalman filter formula computes the assimilated state and parameter vector as:
kkf
k
f
kkkf
k
f
kkkkf
k
f
k
a
k
a
k zKy
HKIy
HzKyy
(23)
Assuming that no direct observations are made on the parameters, and only the state is observed, the following formula
is obtained:
kk
f
kkk
f
kkkkf
k
f
k
a
k
a
k zKyHKIyHzKyy
(24)
Using the polynomial chaos expansions of the forecast state and the parameters:
S
i
iif
k
S
i
iif
k
f
k
f
k
yy
1
1
(25)
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 7
the Kalman filter formula is used to determine the polynomial chaos expansion of the assimilated model and
parameters. For this, first insert the polynomial chaos expansions into the filter formula:
1
1
1
1
1kkS
i
iif
k
S
i
iif
k
kkS
i
iia
k
S
i
iia
k
zK
y
HKI
y
(26)
Note that the term with the observations does not depend on the random variables and is therefore associated with only
the first (constant) basis function. By a Galerkin projection, it can be observed that the polynomial chaos coefficients of
the assimilated state and parameters are:
SizKy
HKIy
ikkif
k
if
kkkia
k
ia
k ,,1,1
(27)
If all the observations are made only on the state of the system, then:
SizKyHKIy
ikk
if
kkkia
k
ia
k ,,1,1
(28)
The covariance of the extended state vector is:
T
kyyk
T
kkkkk
y
k
y
k
yy
kkk
kkk
k
k
k HPHH
PHPP
PP
y
yyyP
00,
cov,cov
,covcovcov
(29)
The Kalman gain reads:
11
0
T
k
k
yykkT
k
k
y
T
k
k
yyT
k
k
yykk
T
kkk HPHR
HP
HPHPHR
HPK
(30)
The parameter estimate is then:
kkk
T
k
k
yykk
T
k
k
yk
a
k yHzHPHRHP 1
(31)
In the polynomial chaos framework the covariance matrices yyP and yP can be estimated from the polynomial chaos
expansion of the solution and the parameters. Then the polynomial chaos coefficients of the parameters are adjusted as:
SiyHzHPHRHPi
kkik
T
k
k
yykk
T
k
k
y
i
k
ia
k ,,1,1
1
(32)
Let’s note that the Kalman filter formula is optimal for the linear Gaussian case. For non-Gaussian uncertainties the
Kalman filter formula is sub-optimal, but is still expected to work.
Another possible approach is to apply the filter formula only once, on a vector containing all the observations from
1t to kt :
SiHyzHHPRHP i
i
T
yy
T
y
iia ,,1,1
1
0
(33)
where
TT
N
T
k
T zzzz )()()( 1 (34)
NHHH ,,diag 1 (35)
NRRR ,,diag 1 (36)
yP y ,cov 0 , yP yy cov with TT
N
T
k
T yyyy )()()( 1 (37)
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 8
The original approach will be called the one-time-step-at-a-time EKF approach, and this alternative approach will be
called the whole-set-of-data-at-once EKF approach. The two approaches are equivalent only for linear systems with
Gaussian assumptions. Even in that case, they might not always be equivalent in practice, due to numerical issues.
The polynomial chaos theory allows for nonlinear propagation of the covariance matrix, which is likely to lead
to improvements over the traditional EKF. The traditional EKF performs a linear propagation of the covariance through
linearized dynamic systems. For example, consider the system 2-1' yy , which has a solution )tanh()( ttyref . Let us
go through one step of filtering from 10 t to 1ft assuming an initial guess 7.00 y with a standard deviation
1.0stdy and a Beta (2, 2) distribution, which would be based on previous filtering steps. The reference solution at
10 t is actually 7616.0)1tanh()( 0 tyref . The observation at 1ft will be 7616.0)1tanh()( fref ty , with an
added measurement noise assumed to be Gaussian with a zero mean and a variance 1% of the value of the observation.
Using 1000 runs in order to account for the noise, the average estimate obtained at 1ft using the polynomial chaos
based EKF estimation method with five terms in the polynomial chaos expansions and 10 collocation points is 0.8124,
while the average estimate obtained with the traditional EKF using linear propagation is 0.8623. In other words, the
polynomial chaos based EKF yields an average error of 0.0508 while the traditional EKF with linear propagation yields
an average error of 0.1007. The collocation approach is explained in section 3.2.
Figure 1(a) shows the distribution of the forecast state obtained with traditional EKF using linear
propagation with 1000 runs. Figure 1(b) shows the distribution of the forecast state obtained with the polynomial
chaos based EKF with 1000 runs. With the polynomial chaos based EKF estimation method, the skewness of the
forecast state can be represented, while using the traditional EKF with linear propagation results in a Gaussian
distribution of the forecast state. . As a consequence, the polynomial chaos based EKF approach leads to a better
estimate, which is obtained using the assimilated state , as shown in Figs. 1(c) and 1(d). For this example, the error obtained using the polynomial chaos based EKF approach is about half the error obtained using the traditional EKF
approach.
(a) (b)
(c) (d)
Fig. 1 Polynomial Chaos Based EKF vs. Traditional EKF Using Linear Propagation Histograms: (a) Forecast
State for EKF with Linear Propagation; (b) Forecast State for Polynomial Chaos Based EKF; (c) Assimilated
State for EKF with Linear Propagation; (d) Assimilated State for Polynomial Chaos Based EKF
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 9
3. INSIGHT INTO THE EKF APPROACH USING SIMPLE MECHANICAL SYSTEMS
3.1. Roll Plane Modeling of a Vehicle
The model used to apply the theory presented in this article is based on the four degree of freedom roll plane model of a
vehicle used in [68] with the addition of a mass on the roll bar, as shown in Fig. 2. The difference is that the suspension
dampers and the suspension springs used in this study are nonlinear and that a mass is added on the roll bar, which
represents the driver, the passenger, and other objects in the vehicle. The added mass M and its position CGd away
from the left end of the roll bar are assumed to be uncertain. It is assumed that there is a passenger, and apriori
distribution of the added mass will therefore be centered in the middle of the bar. This added mass will be represented
as a point mass for the sake of simplicity. Measuring the position of the C.G. of the added mass physically is not
straightforward. However, if a well defined road input can be used and sensors are available, these two parameters can
be estimated based on the observed displacements and velocities across the suspensions.
Fig. 2 Four Degree of Freedom Roll Plane Model (adapted from the model used in [68])
The body of the vehicle is represented as a bar of mass m (sprung mass) and length l that has a moment of
inertia I . The unsprung masses, i.e., the masses of each tire/axle combination, are represented by 1tm and 2tm . A mass
is added on the roll bar, which represents the driver and other objects in the vehicle. That added mass is represented as a
point mass of value M situated at a distance CGd from the left extremity of the roll bar.
The motion variables 1x and 2x correspond to the vertical position of each side of the vehicle body, while the
motion variables 1tx and 2tx correspond to the position of the tires.
The inputs to this system are 1y and 2y , which represent the road profile under each wheel.
If x is the relative displacement across the suspension spring with stiffness ik (i = 1, 2), the force across the
suspension spring is given by:
2,1,3
3, ixkxkxF iiKi (38)
If v is the relative velocity across the damper with a damping coefficient ic (i = 1, 2), the force across the damper
is given by:
M
M and d are uncertainCG
M
M and d are uncertain
dCG
CG
M
M and d are uncertainCG
M
M and d are uncertain
dCG
CG
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 10
)10tanh(2.0)( vcvF iCi (39)
For small angles, i.e. for L
xx 12 small, the equations of motion of the system are
0
1)/(2)(
)(2
22112211
1212
2121
tCtCtKtK
C G
xxFxxFxxFxxF
LdMm
Mxxxx
Mm
(40)
mM
LmdMDwith
L
xxdDMD
LmI
DL
mdDMgxxFxxFDLxxFxxFD
L
xx
L
xx
C G
C G
C GtCtKtCtK
STATIC
)2/(0
2
2
cos
122
2
22221111
1212
2211
(41)
111111111 11 tttCtKtt xykxxFxxFxm (42)
2222222222 2 tttCtKtt xykxxFxxFxm (43)
where 2121
and , , , CCKK FFFF are defined in Eqs. (38) and (39).
In these equations, the variables are expressed versus their position at equilibrium (if the added mass M is not in
the middle, then there are static deflections). S TATIC
L
xx
12 is relative to the position of the ground, which is fixed. It
has to be estimated numerically because of the nonlinearities in the system.
The parameters used in this study are shown in Table 1. They are the parameters used in [68], with the addition of
nonlinearities and uncertainties for M and CGd . For the parameters shown in Table 1, the minimum static angle (i.e.,
the angle of the roll bar with respect to a fixed reference on the ground) is - 1.21 degrees and the maximum static angle
is 1.21 degrees, which corresponds to m 032.012 xx . These values are obtained for )1,1(),( 21 and
)1,1(),( 21 , i.e., for the maximum possible value of M with the added mass as far as possible from the center of
the bar.
Table 1. Vehicle Parameters
Parameter Description Value
m Mass of the roll bar 580 kg
1tm , 2tm Mass of the tire/axle 36.26 kg
1c , 2c Damping coefficients 710.70 N s /m
1k , 2k Spring constants – linear component 19,357.2 N/m
3,1k , 3,2k Spring constants – cubic component 100,000 N/m3
l Length of the roll bar 1.524 m
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 11
I Inertia of the roll bar 63.3316 kg m2
1tk , 2tk Tires vertical stiffnesses 96,319.76 N/m
M Added mass
200 kg +/-50%, with
Beta (2, 2) distribution
CGd Distance between the C.G. of the mass
and the left extremity of the roll bar
0.7620 m +/-25%, with
Beta (2, 2) distribution
The uncertainties of 50% and 25% on the values of M and CGd can be represented as:
1,1),50.01( 11 nomMM (44)
1,1),25.01( 22, nomCGCG dd (45)
where nomM and n o mCGd , are the nominal values of the vertical stiffnesses of the tires ( kg200nomM and
m7620.0, nomCGd ).
It is assumed that the probability density functions of the values of M and CGd can be represented with Beta (2, 2)
distributions [1, 2], with uncertainties of +/- 50% and +/- 25%, respectively. The distributions of the uncertainties
related to the values of M and CGd , defined on the interval 1,1 , are represented in Fig. 3. They have the following
Probability Density Functions (PDFs):
2,1,14
3)(
2 iw ii (46)
(a) (b)
Fig. 3 Beta (2, 2) Distribution: (a) For Value of the Mass; (b) For Value of the Position of the C.G. of the Mass
3.2. Collocation Points
The generalized polynomial chaos theory is explained in [1] in which direct stochastic collocation is proposed as a
less expensive alternative to the traditional Galerkin approach. The collocation approach consists of imposing that the
system of equations holds at a given set of collocation points. If the polynomial chaos expansions contain 15 terms for
instance, then at least 15 collocation points are needed in order to have at least 15 equations for 15 unknown polynomial
chaos coefficients. 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.
In this study, the polynomial chaos expansions of all the variables affected by the uncertainties on M and CGd are
modeled by a polynomial chaos expansion using 15 terms as well, and 30 collocation points will be used to derive the
coefficients associated to each of the 15 terms of the different polynomial chaos expansions. The collocation points used
in this study are obtained using an algorithm based on the Halton algorithm [69], which is similar to the Hammersley
algorithm [70]. One of the advantages of the Hammersley/Halton points used in this study is that when the number of
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 12
points is increased, the new set of points still contains all the old points. Therefore, more points should result in a better
approximation. The collocation points for a Beta (2, 2) distribution, which is used in this study, are shown in [71].
The impact of enforcing dynamics at these few collocation points is discussed in reference [1]. In practice, using
collocation with judicious algorithms such as using the Hammersley/Halton points yields very similar results to what is
obtained with Galerkin, when using enough collocation points. Practically, what needs to be done is checking that
adding more terms and more collocation points does not significantly improve the results. Even though the number of points needed in order to obtain satisfactory results is quite dependent on the example used for parameter estimation, a
satisfactory number of points will typically result in a much faster computation time than any Monte-Carlo based
simulation. An example comparing the computational efficiency of a simulation using a polynomial chaos-based
collocation approach with a Monte Carlo-based simulation yielding the exact same accuracy can be found in Table 3 in
[62].
3.3. Experimental Setting – Road Input
In order to assess the efficiency of the polynomial chaos theory for parameter estimation, M and CGd will be estimated
using observations of four motion variables obtained for a given road input: the displacements across the suspensions (
11 txx and 22 txx ), and their corresponding velocities ( 11 txx and 22 txx ). The road profile is shown in Fig. 4,
and the road input is obtained assuming the vehicle has a constant speed of 16 km/h (10 mph). The road profile can be
seen as a long speed bump. The first tire is subjected to a ramp at 0t , and reaches a height of 10 cm (4”) for a
horizontal displacement of 1m, then stays at the same height for 1m, and goes back down to its initial height. The
second tire is subjected to the same kind of input, but with a time delay of 20% and it reaches a maximum height of only
8 cm.
Fig. 4 Road Profile – Speed Bump
The four motion variables are plotted from 0t to 3t seconds using kg 26.223refM and m 6882.0ref
CGd
(i.e., 2326.01 ref
and 3875.02 ref
) and assuming these values can only be measured with a sampling rate of
s .30 .
However, for the proof of concept of the parameter estimation method presented in this paper, we pretend that the
values of M and CGd are not known, the objective being to estimate those values based on the plot of the four motion
variables shown in Fig. 5. Let’s note that three seconds of data correspond to a horizontal displacement of 13.33 meters.
The end of the speed bump occurs at s 675.0t .
The excitation signal is supposed to be perfectly known. In other words, the road profile shown in Fig. 4 is
supposed to be exactly known and the speed of the vehicle is supposed to be exactly 16 km/h at all time, which enables
us to use any desired sampling rate for the input signal. However, only 10 measurement points are used for the output
displacements and velocities (not counting the measurements at 0t , which give no useful information in order to
estimate the unknown parameter).
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 13
(a) (b)
Fig. 5 Observed States - Displacements and Velocities: (a) Measured; (b) For Nominal Values ( 01 , 02 )
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 [72]. The four degree of freedom
roll plane model used in this article is exactly the same than the one used by [73]. The road inputs used in this article
have also been used by [73], who showed it is possible to perform parameter estimation even when using only 10 time points for 3 seconds of data (i.e., a sampling rate of 0.3s).
The measurements shown in Fig. 5(a) are synthetic measurements obtained from a reference simulation with the
reference value of the uncertain parameter 2326.01 ref
and 3875.02 ref
. Parameters estimation is performed
using the EKF approach. In order to work with a realistic set of measurements, a Gaussian measurement noise with zero
mean and 1% variance is added to the observations shown in Fig. 5 (for the relative displacements and velocities) before
performing parameter estimation.
The state of the system at future times depends on the random initial velocity and can be represented by
T
ttdt
tdx
dt
tdx
dt
tdx
dt
tdxtxtxtxtxty
tt
tt
),(),(
),(),(),(),(),(),(),(),(),( 21
2121
2121
(47)
Assuming that only the displacements across the suspensions ( 11 txx and 22 txx ), and their corresponding velocities
( 11 txx and 22 txx ) can be measured, then
0
0
0
0
010100000
000001010
001010000
000000101
H (48)
and the measurements yield
kkkkkkk RtxtyHz ,0,)()( refref (49)
Measurement errors at different times are independent random variables. The measurement noise k is assumed to be
Gaussian with a zero mean and a variance 1% of the value of )(tx . The diagonal elements of the covariance matrix of
the uncertainty associated with the measurements will still be set to at least 1210
when necessary so that 1
kR can
always be computed. Therefore, the covariance of the uncertainty associated with the measurements is
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 14
4
3
2
1
000
000
000
000
k
k
k
k
k
R
R
R
R
R (50)
where
2
1
12
1 01.0,10max kk zR (51)
2
2
12
2 01.0,10max kk zR (52)
2
3
12
3 01.0,10max kk zR (53)
2
4
12
4 01.0,10max kk zR (54)
(a) (b)
(c) (d)
Fig. 6 EKF Estimation (One-Time-Step-at-a-Time) for Speed Bump Input with 10 Time Points (Noise = 1%):
(a) Mass in the Form of PDF; (b) Distance in the Form of PDF; (c) Mass for Each Term Index; (d) Distance for
Each Term Index
The estimated values of 1 and 2 obtained using the one-time-step-at-a-time EKF approach, which are given by the
first terms of the corresponding polynomial chaos expansions, are 0.2240 1 est
and 0.44152 est
, i.e.,
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 15
kg 222.40es tM and m 6779.0es t
C Gd , which seems to be a good estimation considering that only 10
measurement points were used and that there is noise associated to the measurements. The actual values were
2326.01 ref
and 3875.02 ref
, i.e., kg 26.223r efM and m 6882.0r e f
C Gd . The EKF estimations come in the
form of PDFs, as shown in Fig. 6(a) for M , and Fig. 6(b) for CGd . The estimated values and the corresponding
standard deviations at each time step are plotted in Fig. 6(c) for M , and Fig. 6(d) for CGd . Let’s remember that the
standard deviations are given by Eq. (12).
With 100 sample points (i.e., with time steps of 0.03 s instead of 0.3 s) and a noise level of 1%, the estimated values of
1 and 2 obtained using the one-time-step-at-a-time EKF approach are 0.3988 1
est and 0.08342
est , i.e.,
kg 88.239es tM and m 7461.0es t
C Gd . This is illustrated in Fig. 7.
(a) (b)
(c) (d)
Fig. 7 EKF Estimation (One-Time-Step-at-a-Time) for Speed Bump Input with 100 Time Points (Noise = 1%):
(a) Mass in the Form of PDF; (b) Distance in the Form of PDF; (c) Mass for Each Term Index; (d) Distance for
Each Term Index
Figure 7 shows that the one-time-step-at-a-time EKF approach does not work anymore when using a time step of
0.03 s instead of 0.3 s. Figure 8 shows the absolute error for our two estimated parameters, i.e., refest ,1,1 and
r efes t ,2,2 , with respect to the number of time points, and equivalently, the length of the time step, which is
inversely proportional to the number of time points. It can be observed that a long time step is not really desirable,
which one would expect since less information is available for longer time steps. However, a short time step is even less
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 16
desirable. This seems to be counterintuitive since one would expect that more information would yield more accurate
results. The problem is that the EKF can diverge when using a high sampling frequency. When applying the polynomial
chaos theory to the Extended Kalman Filter (EKF), numerical errors can accumulate even faster than in the general case
due to the truncation in the polynomial chaos expansions. It is shown in Appendix that for the simple scalar system
yay ' with 0a (where a is known), the truncations in the polynomial chaos expansions can prevent the
convergence of the covariance of the assimilated state ay . It is also shown in Appendix that the covariance of the error
after assimilation true
tk
a
k
a
k yyE decreases with the time step t when there is no model error (which is the case for
this study), meaning that using means a larger t results in a smaller error (unless the covariance of ay has not
converged yet, which can happen when t is too large). Figures 7(c) and 7(d), which plot the estimated values of the
two parameters +/- their standard deviations, show that the results with a time step of 0.03 s could not be trusted,
because the EKF was diverging. Indeed, the range of values spanned by the estimated values +/- their standard
deviations at time index k does not always include the range of values spanned by the estimated values +/- their
standard deviations at time index 1k . The curves representing the estimated values +/- the standard deviations of the
estimations can decrease and suddenly increase with new observations or vice versa, unlike what was observed in Figures 6(c) and 6(d), where the curves representing the estimated values +/- their standard deviations smoothly
decrease/increase. Therefore, it is judicious to look at the estimated values and their standard deviations at each time
step. When the estimated values +/- their standard deviations display non-monotonous behaviors, it is a sign that the
sampling frequency should be decreased. Sampling below the Nyquist frequency is usually a necessity in order to
prevent the EKF from diverging. In most cases, sampling below the Nyquist frequency does not result in non-
identifiability issues, but it can in a few rare cases, as illustrated in [72].
(a) (b)
Fig. 8 Absolute Error for the Estimated Parameters ξ1 and ξ 2 with the Nonlinear Half-Car Model for the Speed
Bump with Respect to: (a) the Number of Time Points; (b) the Length of the Time Step
Another possible approach is to apply the filter formula only once, on a vector containing all the observations from 1t
to kt . Using this alternative approach, better results are obtained with a Gaussian measurement noise with zero mean
and 1% variance, for 10 time points (Fig. 9) and for 100 time points (Fig. 10). Applying the EKF formula on the whole
set of data at once with 10 time points yields 0.2321 1 est
and 0.39692 est
, i.e., kg 21.223es tM and
m 6864.0es t
C Gd . Applying the EKF formula on the whole set of data at once with 100 time points yields
0.2305 1 est
and 0.37732 est
, i.e., kg 05.223es tM and m 6901.0es t
C Gd . For this particular road input,
applying the filter formula only once, on a vector containing all the observations clearly yields better results, and this
whole-set-of-data-at-once EKF approach still works with a sampling rate of 0.03 s, while the one-time-step-at-a-time
EKF approach was clearly not working.
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 17
As a conclusion, the one-time-step-at-a-time approach creates more numerical errors. If running the algorithm in
real time is required, a compromise would be to specify a rate of updates not too small in order to avoid creating large
numerical errors and process batches of data with a time length corresponding to that rate (i.e., a mixed whole-set-of-
data-at-once / one-time-step-at-a-time approach) rather than using the one-time-step-at-a-time approach.
(a) (b)
Fig. 9 EKF Estimation (Whole-Set-of-Data-at-Once) for Speed Bump Input with 10 Time Points (Noise = 1%):
(a) Mass in the Form of PDF; (b) Distance in the Form of PDF
(a) (b)
Fig. 10 EKF Estimation (Whole-Set-of-Data-at-Once) for Speed Bump Input with 100 Time Points (Noise =
1%): (a) Mass in the Form of PDF; (b) Distance in the Form of PDF
3.4. Experimental Setting – Out of Phase Sine Input signals at 1 Hz
In order to continue assessing the efficiency of the polynomial chaos theory for parameter estimation, the estimations
will now be performed for a 1-Hz harmonic input, with amplitudes of +/- 0.05 m for 1y and 2y . The input signal is
supposed to be exactly known, which enables us to use any desired sampling rate for the input signal. Figure 11 shows
the harmonic inputs that will be used at 1 Hz. The parameters M and CGd will still be estimated using a plot of four
motion variables: the displacements across the suspensions ( 11 txx and 22 txx ), and their corresponding velocities (
11 txx and 22 txx ). A Gaussian measurement noise with zero mean and 1% variance is still added to the
observations.
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 18
Fig. 11 Road input at 1 Hz
Figure 12 shows the results obtained when using the one-time-step-at-a-time EKF approach with 10 time points,
i.e., with a sampling rate of 0.3 s. Figure 12(c) shows that the estimation of the mass should actually not be trusted for
the reasons explained previously. It can also be observed in Fig. 12(a): the PDF contains values above 300 kg, i.e.,
outside the range of the Beta(2,2) distribution, which means the filter has convergence problems.
(a) (b)
(c) (d)
Fig. 12 EKF Estimation (at 1 Hz with 10 Time Points (Noise= 1%): (a) Mass in the Form of PDF; (b) Distance in
the Form of PDF; (c) Mass at Each Time Index; (d) Distance at Each Time Index
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 19
Figure 13 shows the results obtained when using the whole-set-of-data-at-once EKF approach with the same 10
time points. It yields better results for the estimation of the distance, but not for the estimation of the added mass. This
shows that this alternative approach does not necessarily work better for every problem, even though it often yields
better results, as it clearly did with the speed bump used in section 3.3.
Figure 14 shows the results obtained when using the one-time-step-at-a-time EKF approach with 100 time points,
i.e., with a sampling rate of 0.03 s. The filter clearly diverges and the estimations cannot be trusted, which is especially evident for the estimation of the mass.
Figure 15 shows the results obtained when using the whole-set-of-data-at-once EKF approach with the same 100
time points. The estimation of the mass comes with a large standard deviation, but this approach actually yields an
acceptable estimation for the mass: kg 12.221es tM ( 0.2112 1 est
). In this case, the whole-set-of-data-at-once EKF
approach yields better results with 100 time points than with 10 time points. However, the whole-set-of-data-at-once
approach still does not solve all the drawbacks associated with the use of an EKF. It can be observed that the PDF
contains values outside the range of the Beta(2,2) distribution, i.e., below 100 kg or above 300 kg, so the convergence
problems also appear to affect the whole-set-of-data-at-once approach. When the whole-set-of-data-at-once approach
yields a PDF with a large range of possible values, it is not clear how much it can be trusted. As a conclusion, the EKF
estimation obtained when applying the filter formula only once on the whole set of data can sometimes yield much
better results, but not always, so comparing the results to a different approach (e.g., a Bayesian approach) is strongly recommended. Blanchard et al. [71] performed parameter estimation for the same system using linear springs and
dampers and observed that the PDFs obtained for the linear case and the nonlinear case were quite similar, which
indicates that the problems that have been encountered do not seem to come from the nonlinearities in the springs and
dampers.
(a) (b)
Fig. 13 EKF Estimation (Whole-Set-of-Data-at-Once) at 1 Hz with 10 Time Points (Noise= 1%): (a) Mass in the
Form of PDF; (b) Distance in the Form of PDF
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 20
(a) (b)
(c) (d)
Fig. 14 EKF Estimation (One-Time-Step-at-a-Time) at 1 Hz with 100 Time Points (Noise= 1%): (a) Mass in the
Form of PDF; (b) Distance in the Form of PDF; (c) Mass at Each Time Index; (d) Distance at Each Time Index
(a) (b)
Fig. 15 EKF Estimation (Whole-Set-of-Data-at-Once) at 1 Hz with 100 Time Points (Noise= 1%): (a) Mass in
the Form of PDF; (b) Distance in the Form of PDF
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 21
4. SUMMARY AND CONCLUSIONS
This paper proposes a new computational approach for parameter estimation based on the Extended Kalman Filter
(EKF) and the polynomial chaos theory. The error covariances needed by EKF are computed from polynomial chaos
expansions, and the EKF is used to update the polynomial chaos representation of the uncertain states and the uncertain
parameters. The proposed method has several advantages. It benefits from the computational efficiency of the
polynomial chaos approach in the simulation of systems with a small number of uncertain parameters. The filter formula based on the EKF is also computationally inexpensive. Polynomial chaoses offer an accurate representation of
uncertainties and can accommodate non-Gaussian probability distributions. The approach gives more information about
the parameters of interest than a single value: the estimation comes in the form of a polynomial chaos expansion from
which the aposteriori probability density of the estimated parameters can be retrieved.
For illustration we consider a nonlinear four degree of freedom roll plane model of a vehicle, and we estimate the
uncertain mass and the uncertain position of a body added on the roll bar. The apriori uncertainties on the values of the
added mass and its position were assumed to have a Beta (2, 2) distribution. Synthetic observations of the displacements
and velocities across the suspensions are obtained by adding “measurement noise” to the reference simulation results.
Two different inputs were used: a speed bump and a 1-Hz sinusoidal roll.
Two variations of the approach are discussed: the one-time-step-at-a-time EKF approach, in which the Kalman
filter formula is used at each time step in order to update the polynomial chaos expressions of the uncertain states and
the uncertain parameters, and the whole-set-of-data-at-once EKF approach, which consists of applying the filter formula once, on a vector containing all the observations. For linear systems with Gaussian distribution of uncertainty the two
approaches are theoretically equivalent. For the test problem under consideration the one-time-step-at-a-time EKF
approach yields good estimations for lower sampling rates, but the quality of these estimations deteriorates with
increasing the sampling rate. We explain this counter-intuitive behavior via a rigorous error analysis carried out in the
Appendix. The polynomial chaos truncation errors affect the solution at each filter step; more filter steps mean more
information but also more errors. The truncation errors can accumulate at a fast rate, and over-ride the benefits of the
additional information coming from more measurements. To alleviate this effect we discuss a version of the filter that
uses all the information in a single batch. In most cases, the whole-set-of-data-at-once EKF approach yields more
accurate results than the ones obtained with the one-time-step-at-a-time EKF approach. For a few of the input
excitations and sampling frequencies, however, the results are not very accurate; therefore it is recommended to repeat
the estimation with different sampling rates in order to verify the coherence of the results.
Future work will compare the results obtained with EKF approach to results obtained with a polynomial chaos –
based Bayesian approach. We also plan to apply the proposed techniques to identify parameters of real mechanical
systems for which laboratory measurements are available.
ACKNOWLEDGEMENTS
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
1. Sandu, A., Sandu, C., and Ahmadian, M., 2006, “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).
2. Sandu, C., Sandu, A., and Ahmadian, M., 2006, “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, 15(3), pp. 241-262 (22).
3. Tarantola, A., 2004, Inverse Problem Theory and Methods for Model Parameter Estimation, Society for Industrial
and Applied Mathematics, Philadelphia, PA.
4. Bishwal, J.P.N., 2008, Parameter Estimation in Stochastic Differential Equations, Springer, Berlin, Germany.
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 22
5. Aster, R.C., Borchers, B., and Thurber, C.H., 2005, Parameter Estimation and Inverse Problems, Elsevier
Academic Press. Amsterdam, The Netherlands; Boston, MA.
6. Liu, C.S., 2008, “Identifying Time-dependent Damping and Stiffness Functions by a Simple and Yet Accurate
Method”, Journal of Sound and Vibration, 318, pp. 148–165.
7. Araújo, A.L., Mota Soares, C.M., Herskovits, J., Pedersen, P., 2009, “Estimation of Piezoelastic and Viscoelastic
Properties in Laminated Structures”, Composite Structures, 87(2), pp. 168-174.
8. 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, 194(12-16), pp. 1597-1617.
9. 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, 24(1), pp. 115-123.
10. 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, 32(12), pp. 3011-3022.
11. Fathy, H. K., Kang, D., and Stein, J.,L., 2008, “Online Vehicle Mass Estimation Using Recursive Least Squares
and Supervisory Data Extraction”, Proceedings of the 2008 American Control Conference, pp. 1842-1848.
12. Liang , J.W., 2007, “Damping Estimation via Energy-dissipation Method” Journal of Sound and Vibration, 307, pp. 349–364.
13. Oliveto, N.D., Scalia, G., and Oliveto, G., 2008, “Dynamic Identification of Structural Systems with Viscous and
Friction Damping”, Journal of Sound and Vibration, 318, pp. 911–926.
14. Raїssi, T., Ramdani, N., and Candau, Y., 2004, “Set Membership State and Parameter Estimation for Systems
Described by Nonlinear Differential Equations”, Automatica, 40, pp. 1771–1777.
15. 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.
16. Thompson, B., Vladimirov, I., 2005, “Bayesian Parameter Estimation and Prediction in Mean Reverting Stochastic
Diffusion Models”, Nonlinear Analysis, 63, pp. e2367 – e2375.
17. Wang, J., Zabaras, N., 2005, “Using Bayesian Statistics in the Estimation of Heat Source in Radiation”, International Journal of Heat and Mass Transfer, 48, pp. 15-29.
18. Khan, T., and Ramuhalli, P. (2008), “A Recursive Bayesian Estimation Method for Solving Electromagnetic
Nondestructive Evaluation Inverse Problems”, IEEE Transactions on Magnetics, 44(7), pp. 1845-1855.
19. Nocedal, J., and Wright, S.J., 2006, Numerical Optimization, 2, Springer New York.
20. Horst, R., Pardalos, P. M., and Thoai, N. V., 2000, Introduction to Global Optimization, Second Edition, Kluwer
Academic Publishers, Dordrecht, The Netherlands.
21. Floudas, C.A., 2000, Deterministic Global Optimization: Theory, Methods, and Applications, Kluwer Academic
Publishers, Dordrecht, The Netherlands.
22. Horst, R., and Pardalos, P.M., editors, 1995, Handbook of Global Optimization, 1, Kluwer Academic Publishers,
Dordrecht, The Netherlands.
23. Pardalos, P.M., and Romeijn, H.E., 2002, editors, Handbook of Global Optimization, 2, Kluwer Academic
Publishers Dordrecht, The Netherlands. 24. Liberti, L., and Maculan, N., 2006, Global Optimization: From Theory to Implementation, Springer, Berlin,
Germany.
25. Davis, L., 1991, Handbook of Genetic Algorithms, Van Nostrand Reinhold, New York.
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 23
26. Zhang, B.T., 1999, “A Bayesian Framework for Evolutionary Computation”, Proceedings of the 1999 Congress on
Evolutionary Computation, 1, pp. 722-728.
27. Sun, J., Zhang, Q., Tsang, E.P.K. (2005) “DE/EDA: A New Evolutionary Algorithm for Global Optimization”,
Information Sciences, 169(3-4), pp. 249-262.
28. Zhang, Q., Sun, J., Tsang, E., and Ford, J., 2004, “Hybrid Estimation of Distribution Algorithm for Global
Optimization”, Engineering Computations, 21(1), pp. 91-107. 29. 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, 9(2), pp. 192-200.
30. Kalman, R. E., 1960, “A New Approach to Linear Filtering and Prediction Problems”, Transaction of the ASME-
Journal of Basic Engineering, 82, pp. 35-45.
31. Evensen, G., 1992, “Using the Extended Kalman Filter with a Multi-layer Quasi-geostrophic Ocean Model”,
Journal of Geophysical Research, 97(C11), pp. 17905-17924.
32. Evensen, G., 1993, “Open Boundary Conditions for the Extended Kalman Filter with a Quasi-geostrophic Mode”,
Journal of Geophysical Research, 98(C19), pp.16529-16546.
33. 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, 99(C5), pp. 10143–10162.
34. Blanchard, E., Sandu, A., and Sandu, C., 2007, “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.
35. 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 - 48th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, 6, pp. 6005-6013,
Honolulu, HI.
36. Snyder, C., Bengtsson, T., Bickel, P., and Anderson, J., 2008, “Obstacles to High-dimensional Particle Filtering”,
Monthly Weather Review, 136(12), pp. 4629-4640.
37. 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.
38. 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, 194(2), pp. 773–794.
39. Ghanem, R.G., and Spanos, P.D., 2003, Stochastic Finite Elements, Dover Publications Inc, Mineola, NY.
40. Ghanem, R.G., and Spanos, P.D., 1990, “Polynomial Chaos in Stochastic Finite Element”, Journal of Applied
Mechanics, 57, pp. 197-202.
41. Ghanem, R.G., and Spanos, P.D., 1991, “Spectral Stochastic Finite-Element Formulation for Reliability Analysis”,
ASCE Journal of Engineering Mechanics, 117(10), pp. 2351-2372.
42. Ghanem, R.G., and Spanos, P.D., 1993, “A Stochastic Galerkin Expansion for Nonlinear Random Vibration Analysis”, Probabilistic Engineering Mechanics, 8(3), pp. 255-264.
43. Xiu, D., and Karniadakis, G.E., 2002, “The Wiener-Askey Polynomial Chaos for Stochastic Differential
Equations”, Journal on Scientific Computing, 24(2), pp. 619-644.
44. 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, 124, pp. 51-59.
45. Xiu, D., and Karniadakis, G.E., 2002, “Modeling Uncertainty in Steady-state Diffusion Problems via Generalized Polynomial Chaos”, Computer Methods in Applied Mechanics and Engineering, 191, pp. 4927-4928.
46. Xiu, D., and Karniadakis, G.E., 2003, “Modeling Uncertainty in Flow Simulations via Generalized Polynomial
Chaos”, Journal of Computational Physics, 187, pp. 137-167.
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 24
47. 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.
48. 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.
49. 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.
50. Sandu, C., Sandu, A., and Li, L., 2006, “Stochastic Modeling of Terrain Profiles and Soil Parameters”, SAE 2005
Transactions Journal of Commercial Vehicles, 114(2), pp. 211-220.
51. Blanchard, E., Sandu, C., and Sandu, A., 2007, “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.
52. Blanchard, E., Sandu, A., and Sandu, C., 2007, “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.
53. Soize, C. and Ghanem, R., 2005, “Physical Systems with Random Uncertainties: Chaos Representations with
Arbitrary Probability Measure”, SIAM Journal on Scientific Computing, 2005, 26(2), 395-410.
54. Desceliers, C., Ghanem, R., and Soize, C., 2006, “Maximum Likelihood Estimation of Stochastic Chaos
Representations from Experimental Data”, International Journal for Numerical Methods in Engineering, 2006,
66(6), pp. 978-1001.
55. Desceliers, C., Soize, C., and Ghanem, R., 2007, “Identification of Chaos Representations of Elastic Properties of
Random Media Using Experimental Vibration Tests”, Computational Mechanics, 2007, 39(6), pp. 831-838.
56. 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 - 48th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, 6, pp. 6005-6013,
Honolulu, HI.
57. Li, J. and Xiu, D., 2009, “A Generalized Polynomial Chaos Based Ensemble Kalman Filter with High Accuracy”,
228, pp. 5454-5694.
58. Evensen, G., 2003, “The Ensemble Kalman Filter: Theoretical Formulation and Practical Implementation”, Ocean
Dynamics, 53, pp. 343–367.
59. Lenartz, F., Raick, C., Soetaert, K., and Grégoire M., 2007, “Application of an Ensemble Kalman Filter to a 1-D
Coupled Hydrodynamic-Ecosystem Model of the Ligurian Sea”, Journal of Marine Systems, 68, pp. 327–348.
60. Smith, A. H. C.; Monti, A.; Ponci, F.; 2007, “Indirect Measurements via a Polynomial Chaos Observer” IEEE
Transactions on Instrumentation and Measurement, 56(3), pp. 743 – 752.
61. Wan, X. and Karniadakis, G.E., 2006, “Multi-Element Generalized Polynomial Chaos for Arbitrary Probability
Measures,” SIAM Journal on Scientific Computing, 28(3), pp. 901–928.
62. Cheng, H. and Sandu, A., 2009, “Uncertainty Quantification and Apportionment in Air Quality Models.
Environmental Modeling and Software, 24, pp.917–925.
63. Cheng, H. and Sandu, A., 2009, “Efficient Uncertainty Quantification with the Polynomial chaos Method for Stiff
Systems”, Mathematics and Computers in Simulation, doi:10.1016/j.matcom.2009.05.002.
64. Wiener, N., 1938, “The Homogeneous Chaos”, American Journal of Mathematics, 60, pp. 897–936.
65. Askey, R. and Wilson, J., 1985, “Some Basic Hypergeometric Polynomials that Generalize Jacobi Polynomials”,
Memoirs of the American Mathematical Society, 319, pp. 1–55.
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 25
66. Cohn, S. E., 1997, “An Introduction to Estimation Theory”, Journal of the Meteorological Society of Japan, 75(B),
pp. 257-288.
67. Fisher, M., 2002, Assimilation Techniques(5): Approximate Kalman Filters and singular vectors
68. Simon, D.E., 2001, “An Investigation of the Effectiveness of Skyhook Suspensions for Controlling Roll Dynamics
of Sport Utility Vehicles Using Magneto-Rheological Dampers”, Ph.D. thesis, Virginia Tech, Blacksburg, VA.
69. Halton, J. H., Smith, G. B., 1964, “Radical-inverse Quasi-random Point Sequence”, Communications of the ACM,
7(12), pp. 701–702.
70. Hammersley, J. M., 1960, “Monte Carlo Methods for Solving Multivariables Problems”, Annals of the New York
Academy of Sciences, 86, pp. 844–874.
71. Blanchard, E., Sandu, A., and Sandu, C., 2008, “Parameter Estimation for Mechanical Systems Using an Extended
Kalman Filter”, Technical Report CS-TR-08-18, Computer Science Department of Virginia Tech, Blacksburg, VA.
72. Blanchard, E., Sandu, A., and Sandu, C., 2007, “A Polynomial Chaos Based Bayesian Approach for Estimating
Uncertain Parameters of Mechanical Systems – Part I: Theoretical Approach”, Technical Report TR-07-38,
Computer Science Department of Virginia Tech, Blacksburg, VA.
73. Blanchard, E., Sandu, A., and Sandu, C., 2007, “A Polynomial Chaos Based Bayesian Approach for Estimating
Uncertain Parameters of Mechanical Systems – Part II: Applications to Vehicle Systems” , Technical Report TR-
07-39, Computer Science Department of Virginia Tech, Blacksburg, VA.
LIST OF FIGURES
Fig. 1 Polynomial Chaos Based EKF vs. Traditional EKF Using Linear Propagation Histograms: (a) Forecast State
for EKF with Linear Propagation; (b) Forecast State for Polynomial Chaos Based EKF; (c) Assimilated State
for EKF with Linear Propagation; (d) Assimilated State for Polynomial Chaos Based EKF
Fig. 2 Four Degree of Freedom Roll Plane Model (adapted from the model used in [68])
Fig. 3 Beta (2, 2) Distribution: (a) For Value of the Mass; (b) For Value of the Position of the C.G. of the Mass
Fig. 4 Road Profile – Speed Bump
Fig. 5 Observed States - Displacements and Velocities: (a) Measured; (b) For Nominal Values ( 01 , 02 )
Fig. 6 EKF Estimation (One-Time-Step-at-a-Time) for Speed Bump Input with 10 Time Points (Noise = 1%): (a)
Mass in the Form of PDF; (b) Distance in the Form of PDF; (c) Mass for Each Term Index; (d) Distance for
Each Term Index
Fig. 7 EKF Estimation (One-Time-Step-at-a-Time) for Speed Bump Input with 100 Time Points (Noise = 1%): (a)
Mass in the Form of PDF; (b) Distance in the Form of PDF; (c) Mass for Each Term Index; (d) Distance for
Each Term Index
Fig. 8 Absolute Error for the Estimated Parameters ξ1 and ξ 2 with the Nonlinear Half-Car Model for the Speed
Bump with Respect to: (a) the Number of Time Points; (b) the Length of the Time Step
Fig. 9 EKF Estimation (Whole-Set-of-Data-at-Once) for Speed Bump Input with 10 Time Points (Noise = 1%): (a)
Mass in the Form of PDF; (b) Distance in the Form of PDF
Fig. 10 EKF Estimation (Whole-Set-of-Data-at-Once) for Speed Bump Input with 100 Time Points (Noise = 1%): (a) Mass in the Form of PDF; (b) Distance in the Form of PDF
Fig. 11 Road input at 1 Hz
Fig. 12 EKF Estimation (One-Time-Step-at-a-Time) at 1 Hz with 10 Time Points (Noise= 1%): (a) Mass in the Form
of PDF; (b) Distance in the Form of PDF; (c) Mass at Each Time Index; (d) Distance at Each Time Index
Fig. 13 EKF Estimation (Whole-Set-of-Data-at-Once) at 1 Hz with 10 Time Points (Noise= 1%): (a) Mass in the
Form of PDF; (b) Distance in the Form of PDF
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 26
Fig. 14 EKF Estimation (One-Time-Step-at-a-Time) at 1 Hz with 100 Time Points (Noise= 1%): (a) Mass in the
Form of PDF; (b) Distance in the Form of PDF; (c) Mass at Each Time Index; (d) Distance at Each Time
Index
Fig. 15 EKF Estimation (Whole-Set-of-Data-at-Once) at 1 Hz with 100 Time Points (Noise= 1%): (a) Mass in the
Form of PDF; (b) Distance in the Form of PDF
Fig. A1 Covariance of EN after Convergence with no Model Error (Q = 0, B = 0): (a) R = 0.0001, Mu = 0.0005, a =
-1; (b) R = 0.0001, Mu = 0.0050, a = -1
Fig. A2 Covariance of EN after Convergence for R = 0.0001, Mu = 0.0050, a = -1, Q = 0.01, B = 0 (i.e., Model Error,
but with no Bias): (a) Covariance due to Model Errors; (b) Covariance due to Measurement Noise
Fig. A3 Covariance of EN after Convergence for R = 0.0001, Mu = 0.0050, a = -1, Q = 0.01, B = 1; (i.e., Model
Error, but with Bias): (a) Error due to Model Errors; (b) Covariance of Error due to Measurement Noise
LIST OF TABLES
Table 1: Vehicle Parameters
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 27
APPENDIX: EKF ERROR ANALYSIS
The objective of this analysis is to show that the truncations in the polynomial chaos expansions can prevent the
convergence of the covariance of the assimilated state and that the error can decrease with the length of the time step
when there is no model error (which was the case for this study: the EKF approach assumes that the equations of motion
of the system are perfectly known). This analysis will also show that when model errors are present, a nonzero optimal
time step can exist.
Consider the scalar system yay ' with 0a , which is considered to be the true system, with initial condition truey0 .
It has a well-known analytical solution: taeyty 0)( . After k time steps t which are assumed to be constant, the
“true” value of the state variable y is:
truetaktrue
k
tatrue
k yeyey 01
(A1)
Using the notation taeb , it can also be written as:
tr uektr ue
k
tr ue
k ybyby 01 (A2)
Let’s notice that 0a is equivalent to 10 b .
A perturbed model will be used:
yay' , ayy 0)0( (A3)
It is also assumed that the error model )(t is independent Gaussian with mean B ( B is the bias) and covariance Q .
For the sake of simplicity, it will also be assumed that )(t is fixed during each time interval t , i.e., that it takes the
fixed value 1k between time kt and 1kt .
The state y is propagated using the model equations:
a
eyey
ta
k
a
k
taf
k
111 (A4)
where the superscript f stands for forecast and the superscript a stands for assimilated.
The assimilated state at step k , a
ky , is given by:
obs
kk
f
kk
a
k yKyKy )1( , (A5)
In the 1-dimentional case, each matrix becomes a scalar. Here, it will be assumed that all the kR ’s can be replaced by
R ( wherekR is the covariance matrix of the observational error defined in Eq. (16)), which means that the noise
level associated with the measurements is assumed to be constant. It will also be assumed that 1kH , i.e., we can
directly measure y . Therefore, the Kalman gain at step k is
f
k
f
kk
PR
PK
(A6)
where f
kP is the forecast covariance matrix defined in Eq. (17)
and the assimilated state at step k , a
ky , is given by:
obs
kf
k
f
kf
kf
k
a
k yPR
Py
PR
Ry
(A7)
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 28
Using the notation a
b
a
ec
ta 11
, the model equation can be rewritten as
11 k
a
k
f
k cyby (A8)
Let a
kE be the error at step k after assimilation,
true
k
a
k
a
k yyE (A9)
Then, the “forecast” error at step 1k (before assimilation) is given by:
11 k
a
k
f
k cEbE (A10)
Therefore,
BcEbE a
k
f
k 1 (A11)
)(11 BcEEbEE k
a
k
a
k
f
k
f
k (A12)
)(0
22222
11 ,
eduncorrelat
k
a
k
a
kk
a
k
a
k
f
k
f
k BcEEbBcEEbEE
(A13)
which can also be written as
QcPbP a
k
f
k
22
1 (A14)
where f
kP 1 is the forecast variance at step 1k .
The objective of this analysis is to study the effect of the polynomial chaos approximation. Therefore, a term due to
the truncation in the polynomial chaos expansion will be added to the forecast covariance:
QcPbP a
k
f
k
22
1 (A15)
For the sake of simplicity, will be assumed to be a constant. The assumption about the error being constant can be
regarded as a lower bound on the error since the term due to the truncation always has the same sign ( 0 ), which
means that the covariance is underestimated. Therefore, the error will always be at least that number. Indeed, the
average value of a quantity
1
)(i
jiyy is
1yy , (A16)
which means that its covariance can be expressed as
2
2)cov(
i
iyy (A17)
while the covariance of its truncated polynomial chaos expression is
S
i
iPC yy2
2)cov( (A.18)
Therefore, the term due to the truncation in the polynomial chaos expansion is
1
20)cov()cov(
Si
iPC yyy (A.19)
Thus, the effect of the truncation will be to underestimate the covariance. It will be shown later in Eq. (A32) that
overestimating the covariance is not a problem, but underestimating it too much prevents the convergence of the
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 29
covariance. Therefore, the assumption about the truncation error of the covariance being constant can be regarded as a
lower bound on the error.
Using the notation Qc2 yields
a
k
f
k PbP 2
1 (A20)
Let’s note that is a constant for a constant time interval t . An independent Gaussian noise with mean zero and
covariance R is added to the observations:
10
1
1111
k
truek
k
true
kk
true
k
obs
k ybybyy (A21)
The assimilated state at step 1k , a
ky 1 , is given by
obs
kk
f
kk
a
k yKyKy 11111 )1( (A22)
Using the notation a
kk EE (i.e., kE is the error after assimilation at step k ), the error after assimilation at step 1k
is:
true
k
a
k
true
k
a
kk ybyyyE 1111 (A23)
which yields [71]
a
k
k
a
kkkk
PbR
PbcREbRE
2
1
2
1
1
)( (A24)
The assimilated covariance at step 1k , a
kP 1 , is given by:
f
kkk
f
k
a
k PHKPP 11111 (A25)
which yields [71]
a
k
a
ka
kPbR
PbRP
2
2
1
)( (A26)
Using the recurrence for the error and the covariance after assimilation yields the following Jacobian matrix:
2
2
22
111
2
2
11
11
0
)(
RPb
Rb
RPb
cEbRb
RPb
Rb
dP
dP
dE
dP
dP
dE
dE
dE
a
k
a
k
kkk
a
k
a
k
a
k
k
a
k
a
k
k
k
k
(A27)
which yields the conditions for linear stability:
112
RPb
Rba
k
(A28)
which is equivalent to the following two conditions
RbPb a
k )1(2 and RbPb a
k )1(2 (A29)
It means that the only case for which the covariance converges is when:
RbPb a
k )1(2 (A30)
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 30
or equivalently, when
Qa
bRbPb a
k
2
2 1)1(
, taeb (A31)
which also forces the following condition to be true [71],
0 , i.e. Qa
b2
1
(A32)
It has just been shown that the convergence of a
kP , i.e., the convergence of the covariance of a
ky , is affected by the
truncations in the polynomial chaos expansions. Let’s remind that
a
k
f
k PbP 2
1. It means that overestimating the
covariance is not a problem, but underestimating it too much prevents the convergence of the covariance. It can be
explained by looking at Eq. (A6) and seeing that a very large forecast covariance results in a Kalman gain close to 1,
which means that the assimilated value of the state y will be very similar to the observation and the impact of the
previous error will be gone, which can be seen by looking at Eq. (A22). When the forecast covariance is very small, the
Kalman gain will be close to 0, and the assimilated value of the state y will be very similar to its forecast value, which
means that the convergence of the covariance will be slow.
Looking at Eq. (A26), it can be observed that if the covariance convergences, it converges to
2
2222
2
))1((4)1(
b
RbRbRbPconv
(A33)
After convergence, the recurrence relationship (A24) for the error after assimilation becomes:
conv
kconvkkk
PbR
PbcREbRE
2
1
2
11
)( (A34)
i.e.,
111 kkkk EME (A35)
with
convPbR
RbM
2,
convPbR
Rc2
,
conv
conv
PbR
Pb2
2 )( (A36)
If we rename the steps so that the step 0k is the first step, the error NE ( N steps after
a
kP convergences to convP ,
after assimilation) can be written as:
1
1
1
10
N
iiN
iN
iiN
iN
N MMEME (A37)
The fact that 0E is the error at a new step and therefore has a different value does not matter when studying the
convergence of Eq. (A37) since 02
N
conv
N
PbR
RbM
as N .
Therefore, for large values of N , i.e., long after the covariance has converged.
1
1
1
1
N
iiN
iN
iiN
i
N MME (A38)
For i ’s independent Gaussian with mean B and covariance Q , the term
1
1
N
iiN
iM is Gaussian with mean
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 31
M
B
M
MBBM
NN
i
i
11
11
1
(A39)
and covariance
2
21
1
22
1 M
QMQ
N
i
i
(A40)
When N , the term
1
1
N
iiN
iM is Gaussian with mean M
B
1
and covariance
2
2
1 M
Q
. Similarly, for i ’s
independent Gaussian with mean zero and covariance R , the expression
1
1
N
iiN
iM is Gaussian with mean zero and
covariance 2
22
1
1
22
1
1
M
MRMR
NN
i
i
. When N , the term
1
1
N
iiN
iM is Gaussian with mean zero and
covariance 2
2
1 M
R
.
Since the error model and the measurement noise are not correlated, the covariance of NE is the sum of the
covariance of
1
1
N
iiN
iM and the covariance of
1
1
N
iiN
iM . The mean value of NE is also the sum of the mean
values of the different terms in the sum. Therefore, when N , the mean value of NE is
M
B
1
and the covariance
of NE is
2
22
1 M
RQ
, with , and M defined in Eq. (A36). These expressions can expressed in terms of b , R ,
Q , B , a and , which is implicitly in terms of t , R , Q , B , a and since taeb .
Finding a simple analytical expression of the time step t that minimizes the expression of the mean or the
covariance of NE , or even simply the covariance of
1
1
N
iiN
iM or the covariance of
1
1
N
iiN
iM , is not possible
in the general case, as illustrated by [71]. With Mathematica, we cannot find analytical expressions for the b ’s that set
the derivatives of these expressions with respect to b to 0 [71], which would yield an optimal time step
)(1
optnopt bLoga
t when the second derivative with respect to b is positive. However, in the case where there is no
model error ( 0B , 0Q ), it can be shown analytically that the derivative of the total covariance of NE (which is
also equal to the covariance of
1
1
N
iiN
iM in that case) with respect to b is positive, which means that the
covariance increases with taeb , so decreases with t [71].
A numerical example is shown in Fig. A1, where the covariance of NE is plotted for different time steps,
assuming a perfect model ( 0B , 0Q ). The scalar system used for this example is yy ' , i.e., 1a .
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 32
(a) (b)
Fig. A1 Covariance of EN after Convergence with no Model Error (Q = 0, B = 0): (a) R = 0.0001, Mu = 0.0005,
a = -1; (b) R = 0.0001, Mu = 0.0050, a = -1
Therefore, when there is no model error, it can be shown analytically that the error decreases with t , which
means a larger t results in a smaller error. Figure 9 showed the absolute error for our two estimated parameters with
the nonlinear half-car model for the speed bump with respect to the different corresponding time steps. There was no
model error and a Gaussian measurement noise of mean 0 and variance 1% was added to the observation. It could be
observed that for this case study with a perfect model, the error gets worse for small time steps t . The fact that the
error can get larger as the time step is increase too much was due to the fact that with very few observations, the
covariance had not converged yet. For instance, with a time step of 1.5 seconds, only two significant measurements
were available. The error for the case study yy ' , which was plotted in Fig. A1, was calculated assuming the
covariance had already converged.
With numerical examples, it can be shown that a nonzero optimal time step can exist when 0Q , i.e., when there
is a model error. Figure A2 shows an example where the model error has no bias, i.e., with 0B . The covariance of
NE is the sum of the covariance of
1
1
N
iiN
iM , shown in Fig. A2(a), and the covariance of
1
1
N
iiN
iM , shown
in Fig. A2(b). For the example shown in Fig. A2, the covariance of NE is approximately equal to the covariance of
1
1
N
iiN
iM , so it is not displayed in Fig. A2 since it would be impossible to notice any difference with the
covariance of NE . For the example shown in Fig. A2, there is a nonzero optimal time step that minimizes the
covariance of the estimation error. Even though the covariance of
1
1
N
iiN
iM is very small compared with the
covariance
1
1
N
iiN
iM , the fact that 0Q completely changes the shape of the covariance
1
1
N
iiN
iM , and
therefore the shape of the covariance of NE : there is a nonzero optimal time step that minimizes the covariance of the
estimation error (0.015 s in this case).
Figure A3 shows an example where the model error has a bias, i.e., with 0B . With a large bias, it can be
observed that the error increases with the length of the time step.
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 33
(a) (b)
Fig. A2 Covariance of EN after Convergence for R = 0.0001, Mu = 0.0050, a = -1, Q = 0.01, B = 0 (i.e., Model
Error, but with no Bias): (a) Covariance due to Model Errors; (b) Covariance due to Measurement Noise
(a) (b)
Fig. A3 Covariance of EN after Convergence for R = 0.0001, Mu = 0.0050, a = -1, Q = 0.01, B = 1; (i.e., Model
Error, but with Bias): (a) Error due to Model Errors; (b) Covariance of Error due to Measurement Noise
Extension to the case where the truncation error is proportional to the covariance of the model forecast
For the sake of simplicity, was assumed to be a constant. The assumption about the error being constant could be
regarded as a lower bound on the error since the term due to the truncation always has the same sign ( 0 ). The
purpose of this section is to show that in the case where the truncation error is proportional to the covariance of the model forecast, which is more realistic, taking time steps which are too small can also result in numerical errors
increasing at each time step.
This new assumption can be written as
k
a
k
f
k QcPbP
22
1 with f
ktrunck P 1 (A41)
which is equivalent to
A Polynomial Chaos-Based Kalman Filter Approach for Parameter Estimation of Mechanical Systems
Blanchard E., Sandu A., and Sandu C. 1/11/2012 34
k
a
k
f
k PbP
2
1 with f
ktrunck PQc 1
2
(A42)
It can be noted that 0trunc since 0k . Also, using a recurrence relationship, it can be shown that
1
222 )(k
k
trunc
a
kk QcPbQc (A43)
It can be shown that this still yields
k
a
k
kk
a
kkkk
PbR
PbcREbRE
2
1
2
11
)(,
k
a
k
k
a
ka
kPbR
PbRP
2
2
1
)( (A44)
and that the error kE increases with k when 0k .
Assuming 12 trunc , which is true when a significant number of polynomial chaos terms is used,
)( 222 QcPbQc a
ktrunck (A45)
Using this approximation, the error kE increases with k when
0)( 222 QcPbQc a
ktrunc (A46)
which is equivalent to
a
k
truncPbQc
Qc22
2
with taeb ,
a
ec
ta1 (A47)
Using the approximations tab 1 and tc for small t ’s , it can be noticed that the error kE increases with
k when
a
k
a
k
a
k
a
k
truncP
Qt
PaQtPatP
Qt 2
22
2
)()2(
(A48)
For any nonzero truncation error (i.e., for 0 trunc ), it is possible to find a time step t small enough to result in an
increase of the error kE .