IET Control Theory & Applications
Brief Paper
State estimation for networked systems
with a Markov plant in the presence of
missing and quantised measurements
ISSN 1751-8644
Received on 20th March 2015
Revised on 5th December 2015
Accepted on 25th December 2015
doi: 10.1049/iet-cta.2015.0849
www.ietdl.org
Yingjun Niu
1,2
, Wei Dong
1,2
,Yindong Ji
1,2
1
Department of Automation,Tsinghua University, Beijing 10 0084, People’s Republic of China
2
Tsinghua National Laboratory for Information Science andTechnology (TNList), Beijing 100084, People’s Republic of China
E-mail: weidong@mail.tsinghua.edu.cn
Abstract: This study is concerned with the problem of state estimation for networked systems with a Markov plant consid-
ering the measurement uncertainties. The measurements suffer from both the randomly occurring missing phenomenon
and the quantisation effects. Taking into account the statistical knowledge of the quantised measurements, an approx-
imate minimum mean square error estimate algorithms is derived based on Gaussian assumption, which is referred
to as interacting multiple model Monte Carlo (IMMMC) algorithm. A quantised measurement expectation calculated by
Monte Carlo sampling method is embedded into the Kalman filter under the IMM framework. A simulation example is
provided demonstrating that IMMMC is computationally appealing and presents better estimate performance than the
previous algorithms. Moreover, IMMMC has better mode following ability and can clearly distinguish the occurrence of
measurements missing.
1 Introduction
Nowadays, the rapid development of communication technology
has given rise to the generally called networked systems (NSs).
NSs enable more and more efficient exchanges between differ-
ent components of the practical control systems, which comprise
physical plant, sensor, filter, controller etc., even though located
at different places. Owing to the obvious advantages of NSs such
as lower cost, reduced power requirements, convenient installa-
tion and easy maintenance [1], the problem of state estimation for
NSs has been attracting considerably extensive investigation all the
time. However, subject to the constraints of the communication
capacity, NSs often suffer from unreliable transmission in terms
of packet dropouts and signal delays. Besides, the measurements
are also usually subject to the randomly occurring missing in the
case of the sensor fault or external disturbance [2–5], where only
noise is regarded as the measurements to be transmitted. However,
it is often difficult to obtain the prior knowledge whether the true
signals are contained. Moreover, the limited communication band-
width generally requires the original measurements to be quantised
before transmission. These measurement uncertainties additively
bring complexity and challenges to state estimation over networks.
So far, note that there have been a large number of results
on state estimation for NSs with unreliable network transmission
[3, 5–10], whereas compared with transmission uncertainties, litera-
tures on the measurement uncertainties of missing and quantisation
distortion are relatively insufficient, especially the simultaneous
occurrence of both.
Among the methods of state estimation with quantised mea-
surements, the earliest inspiring results may date back to 1970s,
see [11, 12], wherein a Gaussian-fit recursive algorithm is derived
to approximately deliver the conditional expectation close to the
optimal non-linear estimate. In [13], a moving horizon Monte
Carlo approach is adopted to implement the Gaussian-fit algorithm.
Besides the Kalman-type estimators [14–18], a quantised interact-
ing multiple model particle filter (IMMPF) algorithm has been pro-
posed in [19], wherein a particle filter under the interacting multiple
model (IMM) framework is derived. By representing the posterior
probability density conditioned on quantised measurements with
the summation of the weighted particles, the quantised IMMPF can
strikingly approximate the minimum mean square error (MMSE)
estimate. However, the estimate performance and error conver-
gence may deteriorate without sufficient particles. Thus, it may
expect heavier burden on the computational effort. Different from
the above statistical methods based on Gaussian assumption, more
and more results are proposed with a deterministic framework
inspired by a sector bound approach [20] such as least-square
methods [21, 22] H
∞
methods [23, 24]. In [10], a robust H
∞
state estimation approach is proposed for the Markov jump lin-
ear system (MJLS) with mode-dependent quantised measurements,
whereas the discrete mode needs to be known a priori and the
posterior mode probability cannot be estimated.
As to the problem of state estimation with missing measure-
ments, a Bernoulli distribution is usually adopted to describe the
measurement missing behaviour. There have been different types
of estimators based on Kalman filter [25, 26] and H
∞
filter [27–
29]. Another model of the measurements missing phenomenon is a
two-state Markov chain. In [2], an IMM approach-based distributed
filter is derived by describing the switching of system mode and
measurements missing state with two independent Markov chains.
The measurements uncertainties are so inevitable in practical
NSs that it is of great importance to consider both the measure-
ments missing and quantisation effects in state estimation problem.
However, there are still few results on this aspect so far, to the
best of our knowledge. In [25], a recursive filter is proposed based
on minimising the upper bound on error covariance in terms of the
solutions to two Riccati-like difference equations. It is worth point-
ing out that almost all existing results of the NSs state estimation
are concerned with linear system plant. However, many practical
NSs in reality usually bear a hybrid characterisation plant, espe-
cially the typical representative of hybrid systems, MJLSs. MJLSs
are linear systems whose parameters jumping with time can be
modelled by the realisation of a finite state Markov chain. To
date, in NSs state estimation, scarce effort has been made account-
ing for MJLSs plant, not to mention simultaneously incorporating
the presence of the randomly occurring measurements missing and
quantisation effects. Therefore, it motivates us to shorten the gap
by initiating the research on such a challenging issue in this paper.
IET Control Theory Appl., 2016, Vol. 10, Iss. 5, pp. 599–606
© The Institution of Engineering and Technology 2016 599