ARTICLE IN PRESS

Control Engineering Practice 16 (2008) 392–406

www.elsevier.com/locate/conengprac

Decentralized estimation of overflow losses in a hopper dredger

Zs. Lendeka,, R. Babuškaa, J. Braaksmaa, C. de Keizerb
a
Delft Center for Systems and Control, Delft University of Technology Mekelweg 2, 2628 CD Delft, The Netherlands
b
IHC Systems B.V., P.O. Box 41, 3360 AA Sliedrecht, The Netherlands
Received 30 June 2006; accepted 16 May 2007
Available online 5 July 2007

Abstract

The Kalman filter and its nonlinear variants have been widely used for filtering and state estimation. However, models with severe
nonlinearities are not handled well by Kalman filters. Such a case is presented in this paper: the estimation of the overflow losses in a
hopper dredger. The overflow mixture density and flow-rate have to be estimated based on noisy measurements of the total hopper
volume, mass, incoming mixture density and flow-rate. In order to reduce complexity and make the tuning easier, a decomposition of the
nonlinear process model into two simpler subsystems is proposed. A different type of observer is considered for each subsystem—a
particle filter and an unscented Kalman filter. The performance is evaluated for simulated and real-world data and compared with the
centralized setting for four combinations of the particle filter and the unscented Kalman filter. The results indicate that the distributed
observer achieves the same performance as the centralized one, while leading to increased modularity, reduced complexity, lower
computational costs and easier tuning.
r 2007 Elsevier Ltd. All rights reserved.

Keywords: State estimation; Nonlinear systems; Nonlinear observers; Kalman filter; Particle filter

1. Introduction overcomes some of these deficiencies. The estimates

Many problems require the estimation of states and
possibly uncertain parameters based on a dynamic system
model and a sequence of noisy measurements. Dynamic
systems are usually modeled in the state-space framework,
using a state-transition model, which describes the evolu-
tion of states over time and a measurement model, which
relates the measurement to the states. These models can be
deterministic as well as stochastic.
The most well-known and widely used probabilistic
estimation methods are the Kalman filter (KF) and its
extension to nonlinear systems, the extended Kalman filter
(EKF) (Kalman, 1960; Welch & Bishop, 2002). However,
these methods have severe limitations and may become
unstable even for linear processes. The unscented Kalman
filter (UKF), introduced by Julier and Uhlmann (1997),
Corresponding author. Tel.: +31 15 27 88573; fax: +31 15 27 86679.
E-mail addresses: zs.lendek@tudelft.nl (Zs. Lendek),
r.babuska@tudelft.nl (R. Babuška), j.braaksma@tudelft.nl
(J. Braaksma), cdkeizer@ihcsystems.com (C. de Keizer).
obtained by the UKF are in general more accurate, since
the filter does not rely on linearization, but uses directly the
nonlinear state-transition function. Its superior perfor-
mance has been reported in several publications (Hovland
et al., 2005; Li, Zhang, & Ma, 2004; Stenger, Mendonc- a, &
Cipolla, 2001; van der Merwe & Wan, 2003). Though more
accurate and reliable than the EKF, the UKF still assumes
a unimodal distribution of the states and the handling of
multimodal distributions remains problematic.
Over the last years, particle filters (PFs) (Arulampalam,
Maskell, Gordon, & Clapp, 2002; Doucet, Godsill, &
Andrieu, 2000) have been extensively studied. These filters
have been successfully applied to state-estimation problems,
mainly in the field of target tracking (Hue, Le Cadre, &
Perez, 2002; Li et al., 2007; Nait-Charif & McKenna, 2004;
Sullivan, Blake, Isard, & MacCormick, 2001). The basic idea
behind this technique is to represent probability densities by
a set of samples. In this way, a wide range of probability
densities can be represented, allowing the handling of
nonlinear, non-Gaussian dynamic systems. However, this
representation comes with a higher computational cost,

0967-0661/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2007.05.004
 ARTICLE IN PRESS
Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406
which may render the filter unusable for on-line or real-time
estimation.
Since the above-mentioned methods are suboptimal,
their performance varies, depending on the application
considered. While for a highly nonlinear and non-Gaussian
model, a PF is likely the best option, UKF may also yield
good performance with considerably lower computational
costs. However, the design of an observer for a complex
nonlinear system for on-line estimation is problematic due
to tuning difficulties and large computational costs.
Decentralized estimation has been studied in the context
of large-scale processes and distributed systems. The
architecture in general takes the form of a network of
sensor nodes, each with its own processing facility. In case
of a fully decentralized system, computation is performed
locally and communication occurs between any two nodes.
Each node shares information with other nodes and
computes a local estimate. Computation and communica-
tion is distributed over the network so that a global
estimate can be computed. Several topologies have
been proposed, depending on the particular application.
In case of large-scale processes (Vadigepalli & Doyle,
2003a, 2003b), the network is in general in a hierarchical
form, with several intermediate and one final fusion node.
For distributed systems, such as multiagent societies
(López-Orozco, de la Cruz, Besada, & Ruipérez, 2000;
Roumeliotis & Bekey, 2002; Schmitt, Hanek, Beetz,
Buck, & Radig, 2002), several fusion nodes exist, which
process the data and send the information to the rest
of the nodes. Observers for distributed estimation include,
but are not limited to decentralized KF and EKF filter
(Durrant-Whyte, Rao, & Hu, 1990), information filter and
PFs (Bolic, Djuric, & Hong, 2004; Coates, 2004).
In this paper, it is proposed to decompose a nonlinear
system model into cascaded subsystems, with the possibi-
lity of using different estimation methods for the sub-
systems. Many nonlinear systems can be represented as
cascaded, observable subsystems, which alone are less
complex than the original system. Separate observers can
be designed for the individual subsystems. Moreover,
different types of observers can be combined, depending
on the complexity and nonlinearity of the subsystems. This
setting can be regarded as a cooperative multiagent system.
Each agent has the task of observing one of the subsystems,
possibly using different methods and relying on its own
measurement and the information gathered from other
agents. In turn, each agent communicates its own results to
other agents.
The proposed distributed observer design is applied to
the estimation of the overflow losses in a hopper dredger.
The estimation of overflow losses is an essential step
toward the optimization of the separation process in the
hopper, which is of vital importance for future improve-
ment in dredging efficiency, accuracy and from the
viewpoint of labor saving. In the considered process,
the measured variables are heavily corrupted by noise. The
system is highly nonlinear, and for global state estimation a
393
PF would be required. However, the model can be
represented as two cascaded subsystems, which allows the
use of two observers. For these observers the combinations
of UKF and PF are considered and the four possible
combinations in the distributed setting are compared with
the performance of a centralized PF for the whole system,
both on simulated and experimental data.
The structure of the paper is as follows. Section 2
reviews the UKF and the PF methodology. In Section 3,
the proposed cascaded observer setting is given, while
Section 4 presents the dynamic sedimentation model and
the models used for estimation purposes. Sections 5 and 6
give the results for the simulated and experimental data,
respectively. Finally, Section 7 concludes the paper.
2. Estimation methods
In this section, two methods for estimating the states and
parameters of a nonlinear system are presented. Consider
the following discrete-time, possibly time-varying, non-
linear system:
xk ¼ f ðxk1 ; vk1 Þ, (1)
yk ¼ hðxk ; Zk Þ, (2)
where k is the current time step, x the state variables, v; Z
the noises of known distributions, y the measurements, f
the state transition model, h the measurement model.
Note that the functions f and h may also depend on other
known inputs or parameters. However, for the ease of
notation, these variables are omitted. It is assumed that
system (1)–(2) is observable, in order to be able to estimate
the states.
The goal is to estimate the states of interest. Two methods
are considered: the UKF and the PF. Both methods use
notions from probability theory, however, the UKF is a
deterministic method, while PFs are stochastic. Both filters
are recursive algorithms, that use all the provided informa-
tion (model and observations) to estimate the current state of
the system. The filters work in two steps: prediction and
update. The prediction step uses the system model and the
information incorporated so far in order to predict the
process’ states. This stage is also known as the time update
step, as it projects the current state forward in time. The
update stage uses the latest measurement to modify (correct)
the projected state. This stage is also known as measurement
update, since it incorporates the information brought by the
new measurement.
2.1. Unscented Kalman filter
For linear systems corrupted by white Gaussian noise,
the KF is proven to be an optimal filter in the least mean
square sense. For nonlinear systems, several extensions
exist: the EKF (based on linearizing the models around the
current states), and the family of sigma point KFs (Julier &
Uhlmann, 2002b; van der Merwe, 2004).
 ARTICLE IN PRESS
394 Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406
The fundamental problem of the EKF is that the
approximations will only be valid if all the higher order
derivatives of the nonlinear functions are effectively zero
around the current estimate. Another variant of the KF for
nonlinear systems, the UKF was developed by Julier and
Uhlmann (1997). Since the original formulation of the
UKF, a number of sigma-point filters were described,
which are based on approximating the distribution of the
states by deterministically chosen samples (sigma points)
and differ mostly on how these samples are generated.
These filters preserve the normal distribution and are valid
up to the second order of the Taylor expansion of the
nonlinear functions.
The UKF is based on the unscented transformation,
which computes the statistics of a random variable
undergoing a nonlinear transformation. To compute the
statistics, a number of weighted samples called ‘‘sigma
points’’ are chosen deterministically, so that they comple-
tely capture the mean and covariance of the random
variable.
2.1.1. Unscented transformation
Assume that an nx -dimensional random variable x has to
be propagated through the nonlinear function g in order to
generate y:
y ¼ gðxÞ.
Assume also that x has a known mean xm and a known
covariance Px . In this case, 2nx þ 1 sigma points are
generated deterministically, so that they capture the mean
and variance. One of the selection procedures (Julier &
Uhlmann, 2002a) computes the sigma points Xi as follows:
X0 ¼ x m ; w0 ¼ k=ðnx þ kÞ,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xi ¼ xm þ ½ ðnx þ kÞPx Ti ,
wi ¼ 1=½2ðnx þ kÞ; i ¼ 1; 2; . . . ; nx ,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xi ¼ xm  ½ ðnx þ kÞPx Ti ,
wi ¼ 1=½2ðnx þ kÞ; i ¼ nx þ 1; . . . ; 2nx ,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where k is a scaling parameter, ½ ðnx þ kÞPx i is the ith
row of the matrix square root of ðnx þ kÞPx , and wi is the
weight associated with the ith sample.
The sigma points are now propagated though the
nonlinear function g : Yi ¼ gðXi Þ, i ¼ 0; 1; . . . ; 2nx . The
mean and covariance of y are estimated as
X
2nx
ym ¼ wi Yi ,
i¼0
X
2nx
T
Py ¼ wi ðYi  ym ÞðYi  ym Þ .
i¼0
These estimates are accurate to the second order of the
Taylor series expansion of gðxÞ, for any nonlinear function.
However, in certain cases the computed covariance matrix
can be non-positive semidefinite, in which case the filter
collapses.
2.1.2. UKF algorithm
In order to apply the KF to system (1)–(2), the state
variables are augmented with the state transition and
measurement noise, and the state covariance with the state
transition and measurement covariance. Assuming that the
noises have means v and Z, and covariances Q and R,
respectively, the augmented variables can be expressed as
xak1 ¼ ½xTk1 vTk1 ZTk1 T ; na ¼ dimðxak1 Þ,
Pak1 ¼ diagð½Pxk1 Qk1 Rk1 Þ. (6)
The sigma points are computed based on the augmented
state and covariance.
The prediction is extended with respect to the KF, since
the sigma points have to be computed and propagated
through the state transition model to predict the new states.
The predicted states also have to be propagated through
the measurement model in order to predict the measure-
ment. The equations are described below.
Generate the sigma points for (6), according to (3). The
parameter k provides an extra degree of freedom to ‘‘fine
tune’’ the higher order moments of the approximation, and
can be used to reduce the overall prediction error. When
the states are assumed Gaussian, a useful heuristic is to
select na þ k ¼ 3 (Julier & Uhlmann, 2002a). For other
distributions, a different choice of k might be more
appropriate. In this paper, it is assumed that the states
are Gaussian and the above guideline is used.
Since the augmented state from which the sigma points
are generated has the form (6), the sigma points also have
the form:
Xai;k1 ¼ ½Xxk1 Xvk1 XZk1 i ; i ¼ 0; . . . ; 2na , (7)
where Xxk1 denote the part of the sigma points that
corresponds to the state, Xvk1 the part that corresponds to
ð3Þ the state transition noise and XZk1 the part that
corresponds to the measurement noise.
Propagate the sigma points corresponding to the state
and state transition noise through the state transition
model:
Xxi;kjk1 ¼ f ðXxi;k1 ; Xvi;k1 Þ; i ¼ 0; . . . ; 2na . (8)
Predict the next state:
X
2na
(4) x^ kjk1 ¼ wi Xxi;kjk1 (9)
i¼0
and covariance:
(5) X
2na
Pxkjk1 ¼ wi ðXxi;kjk1  x^ kjk1 ÞðXxi;kjk1  x^ kjk1 ÞT . (10)
i¼0
Propagate the transformed sigma points Xxi;kjk1 and the
part corresponding to the measurement noise XZi;k1
 ARTICLE IN PRESS
Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406 395

through the measurement model: Remarks.

Yi;kjk1 ¼ hðXxi;kjk1 ; XZi;k1 Þ; i ¼ 0; . . . ; 2na . (11)
 The presented procedure is a general form of the UKF.
Predict the measurement: For special cases, such as additive state transition
X2na and/or measurement noise, the computational complex-
y^ kjk1 ¼ wi Yi;kjk1 (12) ity can be reduced (Julier & Uhlmann, 1997).
i¼0  The UKF is not restricted to Gaussian noises, though its
and its covariance: best performance is achieved when the random variables
X
2na are Gaussian.
Pykjk1 ¼ wi ðYi;kjk1  y^ kjk1 ÞðYi;kjk1  y^ kjk1 ÞT . (13)  The UKF is a rather general solution for nonlinear state
i¼0 estimation, but it cannot be used successfully in all
Compute the cross-correlation matrix: situations. The filter may collapse due to the lack of
X
2na robustness: in the case of model-plant mismatch the
Pxy
kjk1 ¼ wi ðXxi;kjk1  x^ kjk1 ÞðYi;kjk1  y^ kjk1 ÞT . (14) estimated posterior covariance can increase in an
i¼0 unbounded fashion.
The update stage remains the same as in the KF.
Compute the Kalman gain: 2.2. Particle filters
Kk ¼ Pxy y 1
kjk1 ðPkjk1 Þ . (15)
KFs represent the distribution of random variables by
Correct the predicted state: their mean and covariance. However, for arbitrary
x^ k ¼ x^ kjk1 þ K k ðyk  y^ kjk1 Þ. (16) distributions or nonlinear processes, this representation is
not sufficient for a reliable estimation and there is no
Correct the covariance: general method to compute the resulting distribution
analytically. Therefore, these methods may become un-
Pxk ¼ Pxkjk1 þ K k Pykjk1 K Tk . (17)
stable for highly nonlinear processes. This is why the PFs
The generic UKF algorithm is given in Algorithm 1. approximate the distributions by samples, which can be

Algorithm 1. Unscented Kalman filter

Input: u; y; Q; R; f ; h; P0 ; x0 ; v0 ; Z0
Output: x; P
for k ¼ 1; 2; . . . do . each sample
Prediction:
xak1 ¼ ½xTk1 vTk1 ZTk1 T ; na ¼ dimðxak1 Þ . augment states
Pak1 ¼ diag½Pxk1 Qk1 Rk1  . augment covariance
Xa0;k1 ¼ xak1
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xai;k1 ¼ xak1 þ ½ ðna þ kÞPak1 Ti ; i ¼ 1; 2; . . . ; na
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xai;k1 ¼ xak1  ½ ðna þ kÞPak1 Ti ; i ¼ na þ 1; . . . ; 2na . compute sigma points

Xxi;kjk1 ¼ f ðXxi;k1 ; Xvi;k1 Þ; i ¼ 0; . . . ; 2na . propagate sigma points

P a
x^ kjk1 ¼ 2n x
i¼0 wi Xi;kjk1
. predict next state
P
Pxkjk1 ¼ 2n a x
i¼0 wi ðXi;kjk1  x ^ kjk1 ÞðXxi;kjk1  x^ kjk1 ÞT . predict covariance

Yi;kjk1 ¼ hðXxi;kjk1 ; XZi;k1 Þ; i ¼ 0; . . . ; 2na . propagate transformed sigma points

P a
y^ kjk1 ¼ 2n wi Yi;kjk1 . predict measurement
y Pi¼0
2na
Pkjk1 ¼ i¼0 wi ðYi;kjk1  y^ kjk1 ÞðYi;kjk1  y^ kjk1 ÞT . predict measurement covariance
P2na
Pxykjk1 ¼
x
i¼0 wi ðXi;kjk1  x ^ kjk1 ÞðYi;kjk1  y^ kjk1 ÞT . cross-correlation matrix

Update:
K k ¼ Pxy y
kjk1 ðPkjk1 Þ
1 . Kalman gain
x^ k ¼ x^ kjk1 þ K k ðyk  y^ kjk1 Þ . correct the state
Pxk ¼ Pxkjk1 þ K k Pykjk1 K Tk . correct the covariance
end for
 ARTICLE IN PRESS
396 Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406

easily computed with, rather than by a compact parametric sample size:

form. 1
The PF uses probabilistic models for the state transition N eff ¼ PN i 2
.
function and the measurement function, respectively: i¼1 ðwk Þ

If N eff drops below a specified threshold N T 2 ½1; N,

pðxk jxk1 Þ; pðyk jxk Þ.
particles are resampled by using Algorithm 3.
The objective is to recursively construct the posterior
Algorithm 2. Particle filter
PDF pðxk jyk Þ of the state, given the measured output yk
and assuming conditional independence of the measure- Input: pðxk jxk1 Þ, pðyk jxk Þ, pðx0 Þ, N, N T
ment sequence, given the states. The PF works in two Initialize:
stages: for i ¼ 1; 2; . . . ; N do
Draw a new particle: xi1 pðx0 Þ
(1) The prediction stage uses the state-transition model to Assign weight: wi1 ¼ N1
predict the state PDF one step ahead. The PDF end for
obtained is called the prior. At every time step k ¼ 2; 3; . . .
(2) The update stage uses the current measurement to for i ¼ 1; 2; . . . ; N do
correct the prior via the Bayes rule. The PDF obtained Draw a particle from importance distribution:
after the update is called the posterior. xik pðxik jxik1 Þ
Use the measured yk to update the weight:
PFs represent the PDF by N random samples (particles) xik w~ ik ¼ wik1 pðyk jxik Þ
with
PN their associated weights wik , normalized so that end for
i w~ i
i¼1 wk ¼ 1. At time instant k, the posterior obtained in Normalize weights: wik ¼ PNk
the previous step, pðxk1 jyk1 Þ, is represented by N samples j¼1
w~ jk
xik1 and the corresponding weights wik1 . To approximate if PN 1 oN T then
ðwik Þ2
the posterior pðxk jyk Þ, new samples xik and weights wik are i¼1

generated. Samples xik are drawn from an (chosen) Resample using Algorithm 3.
importance density function qðxik jxik1 ; yk Þ, and the weights end if
are updated, using the current measurement yk
pðyk jxik Þ pðxik jxik1 Þ Algorithm 3. Resampling
w~ ik ¼ wik1 (18)
qðxik jxik1 ; yk Þ
Input: fðxi ; wi ÞgN
i¼1
and normalized
Output: fðxinew ; winew ÞgNi¼1
w~ i for i ¼ 1; 2; . . . ; N do
P
wik ¼ PN k .
~ jk Compute cumulative sum of weights: wic ¼ ij¼1 wjk
j¼1 w
end for
If the importance density qðxk jxk1 ; yk Þ is chosen equal to Draw u1 from Uð0; N1 Þ
the state-transition PDF pðxk jxk1 Þ, the weight update
for i ¼ 1; 2; . . . ; N do
equation (18) becomes
Find xþi , the first sample for which wic Xui .
w~ ik ¼ wik1 pðyk jxik Þ. Replace particle i: xinew ¼ xþi , winew ¼ N1
The use of the transition prior as the importance density is uiþ1 ¼ ui þ N1
a common choice (Arulampalam et al., 2002) and it has the end for
advantage that it can be easily sampled and the weights are
easily evaluated.
The posterior PDF is represented by the set of weighted The state estimate is computed as the weighted mean of
samples, conventionally denoted by the particles:

X
N X
N

pðxk jyk Þ  wik dðxk  xik Þ, x^ k ¼ wik xik .

i¼1
i¼1
For more details on PFs, refer to Doucet et al. (2000),
where d is the Dirac delta measure.
Arulampalam et al. (2002).
The PF algorithm is summarized in Algorithm 2.
A common problem of PF is the particle degeneracy: Remarks.
after several iterations, all but one particle will have
negligible weights. Therefore, particles must be resampled.  PFs have an important advantage over KFs: they can
A standard measure of the degeneracy is the effective handle not only highly nonlinear processes, but also
 ARTICLE IN PRESS
Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406
arbitrary distributions. In such case, however, the mean
of the posterior cannot be considered a correct estimate
of the state, since no guarantees exist that the posterior
is unimodal.
 Due to the approximation of the posterior with weighted
samples, a large number of samples are necessary for
good performance. Hence, the algorithm is computa-
tionally involved, and not suitable for fast processes.
 While in theory PFs are the best of the above estimators,
for nonlinear systems, they cannot be considered a
universal solution. Mild nonlinearities or fast processes
can be better handled by a KF.
3. Centralized and cascaded observers
Consider a general, observable nonlinear system, de-
scribed as
x_ 1 ¼ f 1 ðx; uÞ; y1 ¼ h1 ðx; uÞ,
..
x_ 2 ¼ f 2 ðx; uÞ; . such a decomposition and the use of cascaded observers
..
. ym ¼ hm ðx; uÞ,
x_ n ¼ f n ðx; uÞ
for which an observer has to be designed. For a large number
of states, and nonlinear equations, the design of an observer
is clearly problematic. If the states and/or measurements are
also corrupted by noise and one uses a PF, then a very large
sample set is needed, in which case the computational costs
may render the observer unusable for on-line estimation.
A solution is to decompose such a system, and design
separate observers for each subsystem, allowing that one
subsystem may use the estimation results of another one.
Such a decomposition is presented in Fig. 1. In this setting,
‘‘centralized observer’’ denotes an observer designed for
system (19) as a whole, while distributed/cascaded ob-
servers are designed for the subsystems.
First, one has to determine under which conditions such
a decomposition is possible. A condition for the decom-
position of the observers is that system (19) can be
partitioned as
x_ 1 ¼ f 1 ðx1 ; uÞ,
x_ 2 ¼ f 2 ðx1 ; x2 ; uÞ,
y1 ¼ h1 ðx1 ; uÞ,
y2 ¼ h2 ðx1 ; x2 ; uÞ
∧
u x1
Observer 1
y1
Observer 2
u The efficiency of the sedimentation process heavily
y2
Fig. 1. Cascaded observers.
397
so that the subsystem
x_ 1 ¼ f 1 ðx1 ; uÞ,
y1 ¼ h1 ðx1 ; uÞ (21)
is observable.
Since both systems (19) and (21) are observable, this also
means that the subsystem
x_ 2 ¼ f 2 ðx1 ; x2 ; uÞ,
y2 ¼ h2 ðx1 ; x2 ; uÞ (22)
is observable. In fact, for the subsystem (22), x1 can be
considered as input.
Assuming, without the loss of generality, that m ¼ 2,
observers may be designed for the two parts separately. At
this point, a comparison of the performance of cascaded
observers and a centralized observer for system (19) does
not exist in the literature. However, several advantages of
can already be recognized: the design and tuning of the
sub-observers is simpler, computational costs can be
(19) reduced, different observers can be used for the subsys-
tems. These properties are illustrated in the application
example presented in the sequel.
4. Estimation of overflow losses in a hopper dredger
The PF and the UKF are applied in the cascaded setting
(20) to the estimation of the overflow losses in a hopper
dredger. Information on the amount of overflow losses is
essential both for decision support and automatic control.
Unfortunately, these losses cannot be reliably measured, due
to the presence of air in the overflow pipe. However, as shown
in this paper, they can be estimated by using mathematical
models and the available on-line measurements.
Before stating the estimation problem, the principle of
the dredging process is briefly explained. The dredger uses
a drag head to excavate soil from the sea bottom. A
mixture of soil and water is transported through a pipe to
the hopper, which is a large cargo hold inside the ship (see
Fig. 2).
The soil gradually settles at the bottom of the hopper,
while excessive water (in fact low-density mixture) is
discharged through an overflow pipe whose level can be
adjusted. As the height of the settled sand layer rises, so
does the concentration of the overflow mixture and
(20) eventually the overflow losses become so high that it is
no longer economical to continue dredging. The ship then
sails back to deliver the load. After the sand is discharged,
the ship sails again to the dredging location and the whole
∧
x2 cycle repeats.
depends on the type of soil and is influenced by the flow-
rate and density of the incoming mixture and the manner
the overflow pipe is controlled. An important factor in the
 ARTICLE IN PRESS
398 Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406

Incoming mixture
Overflow pipe
Qi , ρi
Overflow
Q o , ρo

Hopper
Pump ht ρm
ho

Sand bed ρs hs

Drag head Fig. 3. The sedimentation process in the hopper.

Fig. 2. Schematic drawing of a hopper dredger.

optimization of the dredging performance is the minimiza- overflow pipe is full (e.g., because the valve inside the pipe
tion of the overflow losses. In the literature, a number of is engaged), the following model must be used:
sedimentation models have been proposed (Camp, 1946; pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
van Rhee, 2002), however, these models cannot be used as Qo ¼ k0o 2g maxðht  ho ; 0Þ. (27)
a basis for control or optimization of the dredging process.
Clearly, there is some uncertainty in the modeling of the
The reason is that they are based on detailed (often PDE)
overflow rate. Moreover, due to the model’s switching
modeling of the physical phenomena and contain too many
nature, it is not straightforward to estimate its parameters.
uncertain parameters. Therefore, simplified models are
To model the overflow density ro , the density profile in
used, along with advanced signal processing and estimation
the mixture above the sand bed must be described.
techniques.
Generally, this profile can be approximated as a decreasing
function of the height above the sand, but the exact form of
4.1. Dynamic sedimentation model this function is highly uncertain and time varying. In this
paper, the following saturated affine approximation is
The sedimentation process in a hopper dredger can be used:
described by a model with three state variables: the total
mass in the hopper mt , the total volume V t of the mixture ro ¼ maxðrs  kr ðho  hs Þ; rw Þ. (28)
in the hopper and the mass of the sand bed ms (see Fig. 3).
While the first two variables can be derived from on-line The slope kr must be determined at every time instant such
measurements (the ship draught and the total level ht , that the average mixture density rm , computed from the
respectively), the mass of the sand bed is not measurable. mass balance relations, equals to the average of the density
The flow-rate Qi of the incoming mixture and the overflow profile:
height ho are the manipulated inputs and the incoming Z
1 ht
mixture density ri is in this context regarded as a measured rm ¼ maxðrs  kr ðh  hs Þ; rw Þ dh
disturbance. The volume and mass balance equations are hm h s
given by with hm ¼ ht  hs . Solving this constraint for the model
V_ t ¼ Qi  Qo , (23) (28) yields the following equation for the slope:
8
> 2ðrs  rm Þ 1
_ t ¼ Qi ri  Qo ro .
m (24) >
> for rm 4 ðrw þ rs Þ;
< hm 2
The rate of material sedimentation is a function of the kr ¼ 2
>
> ðrs  rw Þ
settling velocity ðfs Þ and the scouring (erosion) effects ðfe Þ >
: otherwise;
2hm ðrm  rw Þ
_ s ¼ fs ðrm Þfe ðQo ; hm Þ.
m (25)
The overflow rate Qo , the density ro and the functions in where the average mixture density is given by
(25) are modeled by using static relationships as detailed mt  ms r ðmt  ms Þ
below. rm ¼ ¼ s .
V t  ms =rs V t rs  ms
If the outgoing mixture freely flows through the overflow
pipe, the flow-rate Qo is given by Validation based on measured data has shown that this
model is not very accurate, but it suffices for the tuning and
Qo ¼ ko maxðht  ho ; 0Þ3=2 , (26)
first evaluation of the PF.
where ko is an uncertain parameter depending on the The settling function fs describes how the rate of
overflow pipe shape and circumference. However, if the sedimentation depends on the undisturbed settling velocity
 ARTICLE IN PRESS
Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406 399

vs and the mixture density: where vk and Zk are zero mean Gaussian noises, with
!b covariances Q and R, respectively. The equivalent prob-
rm  rw rq  rm abilistic model is expressed as
fs ðrm Þ ¼ Ars vs . (29)
rs  rm rq  rw
pðxk jxk1 Þ ¼ Nðxk ; f ðxk1 Þ; QÞ,
The scouring function fe describes the effect of erosion on
the sand bed due to the flow in the mixture (which is pðyk jxk Þ ¼ Nðyk ; hðxk Þ; RÞ.
considered to be equal to the overflow rate in steady state):
!
Using the above estimation model, a centralized PF has
Q2o
fe ðQo ; hm Þ ¼ max 1  ;0 . (30) already been successfully implemented in the data-acquisi-
kc h2m tion and monitoring system of a hopper dredger (Babuška,
The parameters of the entire model have been determined Lendek, Braaksma, & de Keizer, 2006). Here, different
by fitting the outputs of the simulation model to real data observers are compared in terms of their performance. The
from a ship, by using nonlinear least-squares optimization. following cases are considered:

4.2. The estimation problem
In order to estimate at each time step the overflow
density and flow-rate, the volume and mass balance
equations were discretized by using the Euler method:
V t;k ¼ V t;k1 þ TðQi;k1  Qo;k1 Þ,
(1) A centralized observer to simultaneously estimate the
values of both Qo and ro , see Fig. 4.
(2) Cascaded observers: the first observer estimates Qo
based on the volume balance (31), and the second
estimates ro based on the mass balance (32) and the
values obtained for Qo by the first observer, see Fig. 5.
(31)

mt;k ¼ mt;k1 þ TðQi;k1 ri;k1  Qo;k1 ro;k1 Þ, (32) Note, that models (31)–(34) only approximate the
underlying true process. If the data were generated based
where the sampling period is T ¼ 5 s, which is also the on these models, a KF could be used. For the data
sampling period of the on-board data-acquisition system. generated from the sedimentation model, however, the
These state equations are augmented with a random-walk results obtained by the KF are too noisy and the KF
model for Qo and ro : becomes unstable for the experimental data.
Qo;k ¼ Qo;k1 þ Q;k1 , (33) It was found that the UKF cannot simultaneously
estimate both Qo and ro . Therefore, for the centralized
ro;k ¼ ro;k1 þ r;k1 . (34) observer, only PF is considered, while in the cascaded
setting both PF and UKF are used. The observers are first
The motivation for this choice results from the process applied to the simulated data and then, with the same
description in Section 4. The sedimentation models are parameters, to real measured data.
based on empirical modeling of the physical phenomena There is one more setting of observers that was
and contain too many uncertain parameters. By using a considered for this specific application: independent
random walk model, the use of the uncertain overflow observers. In this setting, one uses two observers: the first
model (26)–(27), the settling and scouring functions estimates Qo based on the volume balance (31). The second
(29)–(30), and the uncertain parameters is circumvented. estimates Qo ro based on the mass balance (32). The value
The augmented state, input and output vectors are defined of ro can be computed afterwards by dividing the estimate
as
0 1
Vt
B C ! !
B mt C Qi Vt
x¼B B C Vt
C; u ¼ ri ; y ¼ mt . Qo
@ Qo A Qi
Observer
ro i
o
mt
Measurements are available for the inputs Qi and ri and
the outputs V t and mt . The objective is to estimate Qo and
Fig. 4. Centralized observer.
ro on-line. The corrupting noises are considered zero-mean
Gaussians ðxi;k Nð0; nxi ÞÞ, and their standard deviations
are determined experimentally. Qo
Vt
Note, that in this particular case, the estimation model is Qi
Observer 1
described as Observer 2 o
i
xk ¼ f ðxk1 Þ þ vk1 , mt

yk ¼ hðxk1Þ þ Zk , Fig. 5. Cascaded observers.

 ARTICLE IN PRESS
400 Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406

of Qo ro by the estimate of Qo obtained by the first 5. Results for simulated data

observer. However, when working with experimental data,
the computation of ro means dividing noisy variables and
leads to large errors and therefore the results are not
presented here.
Recall that the model used for simulation is the one
presented in Section 4.1, while the one used for estimation is
V t;k ¼ V t;k1 þ TðQi;k1  Qo;k1 Þ þ V ;k1 ,

mt;k ¼ mt;k1 þ TðQi;k1 ri;k1  Qo;k1 ro;k1 Þ þ m;k1 ,

Table 1
Standard deviations used in the estimation model (35)
Qo;k ¼ Qo;k1 þ Q;k1 ,
Variable State transition Measurement

V t ðm3 Þ 0 10 ro;k ¼ ro;k1 þ r;k1 (35)

mt (kg) 3000 12 000 with V ;k Nð0; nV Þ, m;k Nð0; nm Þ, Q;k Nð0; nQ Þ,
Qo ðm3 =sÞ 0.25 –
r;k Nð0; nr Þ and Nð0; nÞ being a zero-mean, n2 covariance
ro ðkg=m3 Þ 5 –
Gaussian random noise.

11.5 1450

11 1400
10.5 1350
10
1300
9.5
ρo [kg/m3]
Qo [m3/s]

1250
9
1200
8.5
1150
8

7.5 1100

7 1050

6.5 1000
1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6
Time [h] Time [h]

Fig. 6. Centralized observer: results for Qo (a) and ro (b) using the particle filter (solid line—simulated data, dotted line—estimate).

60 50

45
50
40

40 35

30
# samples

# samples

30 25

20
20
15

10
10
5

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −100 −50 0 50 100
Qo [m3/s] ρo [kg/m3]

Fig. 7. Centralized observer: residuals for Qo (a) and ro (b).

 ARTICLE IN PRESS
Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406 401

For this simulation, only the inputs Qi and ri are fed
with experimental data, corrupted by noise. The remaining
variables are computed in simulation without adding noise.
The results obtained with the different configurations of
observers are compared to the simulated values of ro and
Qo . The standard deviations of the state transition and
measurement noise are given in Table 1. The PF used 1000
samples, with resampling at a threshold of N T ¼ 900. The
presented results are the average of 30 simulations.
5.1. Centralized observer
The results obtained with a PF, based on model (35) are
presented in Fig. 6. The maximum standard deviation
computed point-wise for 30 trajectories of the state
estimated by the PF is 0.1 for Qo and 5.1 for ro .
The residuals are computed as the difference between the
simulated and estimated values of Qo and ro , respectively.
The distribution of the residuals is presented in Fig. 7,
while their statistics are given in Table 2.
5.2. Cascaded observers
This setting involves two observers in a cascade. The first
one estimates Qo using the volume balance (31) and a
random walk model for Qo . The second observer estimates
ro based on the mass balance (32), a random walk model
for ro and the result obtained for Qo by the first observer
(see also Fig. 5). Two types of filters are compared: the
UKF and the PF.
The dynamic model is decomposed into two subsystems.
The first observer uses the model
V t;k ¼ V t;k1 þ TðQi;k1  Qo;k1 Þ þ V ;k1 ,
Qo;k ¼ Qo;k1 þ Q;k1 , (36)
where V t is the measured output. The second observer uses
the model
mt;k ¼ mt;k1 þ TðQi;k1 ri;k1  Qo;k1 ro;k1 Þ þ m;k1 ,
ro;k ¼ ro;k1 þ r;k1 , (37)
where mt is the measured output.
The results obtained for Qo are presented in Fig. 8.
The maximum standard deviation computed point-wise
for the 30 Monte Carlo simulations is 0.095. The statistics
of the residuals’ distributions are given in Table 3. These
statistics are comparable with those obtained with the
centralized observer.
For the cascaded observer the following combinations
are considered: UKF for both Qo and ro , PF for both Qo
and ro , UKF for Qo and PF for ro , and PF for Qo and
UKF for ro . In what follows, the observer setting is
denoted as observer 1–observer 2, i.e., PF–PF denotes that
PF is used for both Qo and ro , UKF–PF means that UKF
is used for Qo and PF for ro , etc.

Table 2 Table 3
Statistics of residuals for the centralized observer Statistics of residuals of Qo ðm3 =sÞ

Mean Standard deviation Mean Standard deviation

Qo ðm3 =sÞ 0.0135 0.5863 PF 0.0094 0.5953

ro ðkg=m3 Þ 11.6858 22.6426 UKF 0.0222 0.5582

11.5 11.5

11 11

10.5 10.5

10 10

9.5 9.5
Qo [m3/s]

Qo [m3/s]

9 9

8.5 8.5

8 8

7.5 7.5

7 7

6.5 6.5
1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6
Time [h] Time [h]

Fig. 8. Cascaded observers: results for Qo using PF (a) and UKF (b) (solid line—simulated data, dotted line—estimate).
 ARTICLE IN PRESS
402 Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406

The results obtained for ro using Qo estimated by PF
and UKF are presented in Fig. 9. The maximum point-
wise standard deviation of 30 Monte Carlo simulations
for ro , based on the results of Qo given also by PF is
5.8, while for ro based on the results of Qo given by
UKF is 2.96. The distribution of the residuals is given
in Fig. 10. The statistics of the residuals can be found in
Table 4.
5.3. Discussion
For simulated data, a good estimate is obtained with the
centralized observer (PF) for Qo , but the estimate of ro is
delayed relative to the one simulated (Fig. 6).
For the cascaded observers, considering only the
mean of the residuals (Table 4), one can conclude
that the UKF performs the best for the estimation of ro ,
much better than the centralized observer or the PF.
However, considering the standard deviations (Table 4
and Fig. 10), one can see that the estimates of ro obtained
by UKF are much noisier than those obtained by the
PF, both centralized and cascaded. Comparing the
results obtained by PF for ro with the centralized observer
(Table 2), it can be seen that using a combination of two
PFs leads to approximately the same results as the
centralized filter. However, the best result, based on the
statistics presented in Tables 3 and 4 was obtained with
cascaded observers, the combination UKF for Qo and PF
for ro .
6. Results for experimental data
For the experimental data, the same model and
parameter values are considered as for the simulated

1450 1450

1400 1400

1350 1350

1300 1300
ρo [kg/m3]

ρo [kg/m3]

1250 1250

1200 1200

1150 1150

1100 1100

1050 1050

1000 1000
1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6
Time [h] Time [h]

1450 1450

1400 1400

1350 1350

1300 1300
ρo [kg/m3]

ρo [kg/m3]

1250 1250

1200 1200

1150 1150

1100 1100

1050 1050

1000 1000
1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6
Time [h] Time [h]

Fig. 9. Estimates of ro with the four possible filter combinations (solid line—simulated data, dotted line—estimate). (a) PF–PF. (b) PF–UKF.
(c) UKF–PF. (d) UKF–UKF.
 ARTICLE IN PRESS
Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406 403

50 50

45 45

40 40

35 35

30 30
# samples

# samples
25 25

20 20

15 15

10 10

5 5

0 0
−100 −50 0 50 100 −100 −50 0 50 100
ρo [kg/m3] ρo [kg/m3]

50 50

45 45

40 40

35 35

30 30
# samples

# samples

25 25

20 20

15 15

10 10

5 5

0 0
−100 −50 0 50 100 −100 −50 0 50 100
ρo [kg/m3] ρo [kg/m3]

Fig. 10. Cascaded observers: residuals for the estimates of ro . (a) PF–PF. (b) PF–UKF. (c) UKF–PF. (d) UKF–UKF.

Table 4 computed as
Statistics of residuals of ro ðkg=m3 Þ
1
Qo ro Mean Standard deviation
Qo;k ¼ Qi;k  ðV t;kþ1  V t;k Þ,
T
PF PF 12.1986 25.1174
PF UKF 0.3354 54.3970 Qi;k ri;k  ð1=TÞðmt;kþ1  mt;k Þ
ro;k ¼ . (38)
UKF PF 10.6790 21.5436 Qo;k
UKF UKF 1.2536 54.2936
As the result of this computation is very noisy, a first order
anti-causal low-pass filter was applied to the measured data
before computing Qo and ro , with an experimentally
data and the same combinations of observers. The chosen cut-off frequency of 0.05 Hz.
presented results for the PF are the average of 30
simulations.
Since no measurements of Qo and ro are available, 6.1. Centralized observer
the results are compared to the values of ro and Qo
directly computed from the volume and mass balance, i.e., The results obtained with the centralized observer (PF)
Eqs. (31)–(32). The overflow rate and the density are are presented in Fig. 11. The maximum standard deviation
 ARTICLE IN PRESS
404 Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406

computed point-wise over the 30 Monte Carlo simulations The sample-wise maximum standard deviation of the 30
is 1.76 for Qo and 42.87 for ro . Monte Carlo simulations for ro , based on the results of Qo
given also by PF is 36.91, while for ro based on the results
6.2. Cascaded observers of Qo given by UKF is 36.05.

The results obtained by the UKF and PF for Qo are
presented in Fig. 12. The estimate of Qo obtained by the
UKF is noisier than that obtained by the PF, but
comparable to the result obtained by the centralized
observer. The maximum sample-wise standard deviation
of the Monte Carlo simulations for Qo is 0.55.
For the cascaded setting, the previous four combinations
are considered. The results obtained for ro with the four
combinations are presented in Fig. 13.
6.3. Discussion
For experimental data, a centralized observer (Fig. 11)
obtains a reasonably good estimate for ro , but the estimate
of Qo is noisier than the one computed by (38).
From Fig. 13 it can be seen that the UKF cannot handle
the estimation of ro , probably due to the high noise level of
the measured variables. By using the combinations of
PF–PF and UKF–PF, approximately the same results are

15
1500
14

13 1400

12
1300
11
Qo [m3/s]

ρo [kg/m3]

10 1200
9
1100
8

7
1000
6

5 900
1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6
Time [h] Time [h]

Fig. 11. Centralized observer: results for Qo (a) and ro (b) using the particle filter (solid line—computed by (38), dotted line—estimate).

15 15

14 14

13 13

12 12

11 11
Qo [m3/s]

Qo [m3/s]

10 10

9 9

8 8

7 7

6 6

5 5
1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6
Time [h] Time [h]

Fig. 12. Cascaded observers: results for Qo using PF (a) and UKF (b) (solid line—computed by (38), dotted line—estimate).
 ARTICLE IN PRESS
Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406 405

1500 1500

1400 1400

1300 1300
ρo [kg/m3]

ρo [kg/m3]
1200 1200

1100 1100

1000 1000

900 900
1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6
Time [h] Time [h]

1500 1500

1400 1400

1300 1300
ρo [kg/m3]

ρo [kg/m3]

1200 1200

1100 1100

1000 1000

900 900
1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6
Time [h] Time [h]

Fig. 13. Cascaded observers: estimates of ro using the four possible combinations (solid line—computed by (38), dotted line—estimate). (a) PF–PF.
(b) PF–UKF. (c) UKF–PF. (d) UKF–UKF.

obtained for ro . These results are also comparable to those
obtained by a PF as the centralized observer.
7. Conclusions
This paper proposed a distributed setting for nonlinear
state-estimation. In many real-life applications, a complex
process model can be decomposed into simpler subsystems,
and observers can be designed for these individual
subsystems. This partitioning leads to increased modularity
and reduced complexity of the problem, with reduced
computational costs. Moreover, since the design and
tuning of the observers becomes simpler, the overall
performance can also be improved.
The distributed observer setting was applied to the
estimation of overflow losses in a hopper dredger. The
results were compared with those obtained by the
centralized observer. The overflow losses, represented by
the overflow mixture density are estimated on the basis of
the measured total hopper volume, hopper mass, incoming
mixture density and flow-rate.
The proposed approach uses straightforward nonlinear
mass balance equations and does not rely on complex
overflow and sedimentation models which contain uncer-
tain parameters and empirical functional relationships.
The performance was evaluated in simulations and with
real measurements. The simulation results for this applica-
tion clearly indicate the best combination: cascaded
observers, using UKF or PF for the simpler subsystem
(flow-rate) and PF for the more complex one (density).
In the future research, the theoretical conditions under
which such a distribution of the process and the estimation
is possible while maintaining the same performance as a
centralized observer will be investigated.
 ARTICLE IN PRESS
406 Zs. Lendek et al. / Control Engineering Practice 16 (2008) 392–406
Acknowledgments
This research is in part sponsored by Senter, Ministry of
Economic Affairs of the Netherlands within the projects
Artificial Intelligence for the Control of a Hopper Dredger
(Grant no. TSMA 2017) and the project Interactive
Collaborative Information Systems (Grant no. BSIK03024).
References
Arulampalam, S., Maskell, S., Gordon, N. J., & Clapp, T. (2002). A
tutorial on particle filters for on-line nonlinear/non-Gaussian Bayesian
tracking. IEEE Transactions on Signal Processing, 50, 174–188.
Babuška, R., Lendek, Zs., Braaksma, J., & de Keizer, C. (2006). Particle
filtering for on-line estimation of overflow losses in a hopper dredger.
In 2006 American control conference, Minneapolis, Minnesota, USA
(pp. 5751–5756).
Bolic, M., Djuric, P. M., & Hong, S. (2004). Resampling algorithms and
architectures for distributed particle filters. IEEE Transactions on
Signal Processing, 53(7), 2442–2450.
Camp, T. (1946). Sedimentation and the design of settling tanks.
Transactions on ASCE 895–936.
Coates, M. (2004). Distributed particle filters for sensor networks. In
IPSN’04 (pp. 99–107).
Doucet, A., Godsill, S. J., & Andrieu, C. (2000). On sequential Monte
Carlo sampling methods for Bayesian filtering. Statistics and Comput-
ing, 10, 197–208.
Durrant-Whyte, H. F., Rao, B. Y. S., & Hu, H. (1990). Toward a fully
decentralized architecture for multi-sensor data fusion. In Proceedings
of the IEEE international conference on robotics and automation (Vol. 2,
pp. 1331–1336).
Hovland, G. E., von Hoff, T. P., Gallestey, E. A., Antoine, M., Farruggio,
D., & Paice, A. D. B. (2005). Nonlinear estimation methods for
parameter tracking in power plants. Control Engineering Practice, 13,
1341–1355.
Hue, C., Le Cadre, J., & Perez, P. (2002). Tracking multiple objects with
particle filtering. IEEE Transactions on Aerospace and Electronic
Systems, 38(3), 791–812.
Julier, S. J., & Uhlmann, J. K. (1997). A new extension of the Kalman
filter to nonlinear systems. In Proceedings of 11th international
symposium on aerospace/defense sensing, simulation and controls
(pp. 182–193).
Julier, S. J., & Uhlmann, J. K. (2002a). The scaled unscented
transformation. In Proceedings of the IEEE American control
conference, Anchorage AK, USA (pp. 4555–4559).
Julier, S. J., & Uhlmann, J. K. (2002b). Reduced sigma point filters for the
propagation of means and covariances through nonlinear transforma-
tions. In Proceedings of the 2002 American control conference (Vol. 2,
pp. 887–892).
Kalman, R. E. (1960). A new approach to linear filtering and prediction
problems. Transactions of the ASME—Journal of Basic Engineering,
82, 35–45.
Li, P., Goodall, R., Weston, P., Ling, C. S., Goodman, C., & Roberts, C.
(2007). Estimation of railway vehicle suspension parameters for
condition monitoring. Control Engineering Practice, 15, 43–55.
Li, P., Zhang, T., & Ma, N. (2004). Unscented Kalman filter for visual
curve tracking. Image and Vision Computing, 22, 157–164.
López-Orozco, J. A., de la Cruz, J. M., Besada, E., & Ruipérez, P. (2000).
An asynchronous robust and distributed multisensor fusion system for
mobile robots. International Journal of Robotics Research, 19(12),
914–932.
Nait-Charif, H., & McKenna, S. J. (2004). Tracking poorly modelled
motion using particle filters with iterated likelihood weighting. In The
Asian conference on computer vision systems (ACCV04) (pp. 156–161).
Roumeliotis, S. I., & Bekey, G. A. (2002). Distributed multirobot
localization. IEEE Transactions on Robotics and Automation, 18,
781–795.
Schmitt, T., Hanek, R., Beetz, M., Buck, S., & Radig, B. (2002).
Cooperative probabilistic state estimation for vision-based autono-
mous mobile robots. IEEE Transactions on Robotics and Automation,
18(5), 670–684.
Stenger, B., Mendonc- a, P. R. S., & Cipolla, R. (2001). Model-based hand
tracking using an unscented Kalman filter. In Proceedings of the British
machine vision conference, Manchester, UK (Vol. I, pp. 63–72).
Sullivan, J., Blake, A., Isard, M., & MacCormick, J. (2001). Bayesian
object localisation in images. International Journal of Computer Vision,
44(2), 111–135.
Vadigepalli, R., & Doyle, F. J., III (2003a). A distributed state estimation
and control algorithm for plantwide processes. IEEE Transactions on
Control Systems Technology, 11, 119–127.
Vadigepalli, R., & Doyle, F. J. (2003b). Structural analysis of large-scale
systems for distributed state estimation and control applications.
Control Engineering Practice, 11, 895–905.
van der Merwe, R. (2004). Sigma-point Kalman filters for probabilistic
inference in dynamic state-space models. Ph.D. Thesis, University of
Washington.
van der Merwe, R., & Wan, E. (2003). Gaussian mixture sigma-point
particle filters for sequential probabilistic inference in dynamic state-
space models. In Proceedings of IEEE international conference
on acoustics, speech and signal processing (ICASSP), Hong Kong
(pp. 701–704).
van Rhee, C. (2002). On the sedimentation process in a trailing suction
hopper dredger. Ph.D. Thesis, TU Delft.
Welch, G., & Bishop, G. (2002). An introduction to the Kalman filter.
Technical Report TR 95-041. NC, USA: Department of Computer
Science, University of North Carolina.