Abstract— In this paper, the problem of state estimation for
Markov Jump Linear Systems (MJLSs) with quantized
measurements is investigated. A quantized interacting multiple
model particle filter (IMMPF) algorithm with a combination of
the interacting multiple model (IMM) estimation framework
and a particle filter is proposed. Our proposed algorithm is
efficient to overcome the exponential computing difficulty with
the application of IMM hypothesis merging method and deal
with the nonlinearity of the quantizer by representing the
posterior probability density under quantized observation with
the particle filter. Both the state and the probability distribution
of the mode can be estimated in our estimation framework.
Simulation results demonstrate that the quantized IMMPF has
significant advantage in estimate accuracy over standard IMM
directly using quantized measurements and performs better
than moving horizon Monte Carlo (MHMC) method on time
exhausting aspect.
I. INTRODUCTION
In networked control systems (NCSs) or wireless sensor
networks (WSNs), raw measurements are usually quantized
before transmission due to the limited communication
capability such as the finite bandwidth. With the nonlinearity
of the quantizer, the state estimation becomes a nonlinear and
non-Gaussian estimating problem [1].
An early survey on this subject can be found from Curry’s
result [2, 3] referred to as Gaussian-fit algorithm based on
Gaussian assumption, which in general requires numerical
integration to get the optimal state estimate. In [1], a numerical
method is proposed to implement the approximate minimum
mean square error (MMSE) estimator. In [4, 5], the authors
derive a moving horizon Monte Carlo (MHMC) approach to
estimate the state of linear discrete-time systems. Besides
these Kalman filter based estimators, much of the recent work,
e.g., [6, 7] applies particle filter, which represents posterior
density function under quantized observation by a set of
random samples with associated weights. However, most of
these results are attributed to the linear time invariant systems
(LTISs) and less attention has been paid to hybrid systems.
As a representative of hybrid systems, Markov Jump
Linear Systems (MJLSs) are linear systems whose parameters
evolve with time according to the realization of a finite state
Markov chain [8-10]. MJLSs have been extensively
investigated because of their application in a wide variety of
fields, such as electrical engineering, signal processing, target
tracking and communications. Although many results have
been proposed on the state estimation of MJLSs (see [8, 11-14]
*Resrach supported by the Natural Science Foundation of China under
Grants 61104019, 61374123, and the Tsinghua University Initiative
Scientific Research Program.
Y. Niu, W. Dong and Y. Ji are with the Department of Automation,
Tsinghua University, and Tsinghua National Laboratory for Information
and references therein), methods of the state estimation for
MJLSs with quantized measurements are still scarce. In [15],
the authors extend the MHMC strategy to the MJLSs state
estimation with quantized measurements by taking into
account the statistical knowledge of the quantized intervals
and making full use of the Markov statistics of the mode
jumping to handle the unknown modes. However, more
accurate estimation results may bring larger computational
complexity. In [16] , a robust H∞ state estimation approach is
proposed for the ItÔ stochastic systems with Markov jump
parameters and mode-dependent quantized measurements,
whereas the discrete state estimation can’t be obtained as it
depends on the prior mode information.
In this paper, to deal with the problem of state estimation
for MJLSs with quantized measurements, a type of quantized
interacting multiple model particle filter (IMMPF) algorithm
combing the interacting multiple model (IMM)[13] estimation
framework and a particle filter approach is presented. The
IMM method is the most widely used suboptimal state
estimation approach for MJLSs [17-19]. Based on the
hypothesis pruning and merging techniques, IMM presents
excellent estimate performance with low computational cost.
Meanwhile, due to the nonlinearity of the quantizer, the mean
and covariance updated by the Kalman filter are no longer
sufficient to describe the posterior probability density of the
state. Instead, a particle filter (see [20, 21]), is applied to deal
with the nonlinearity difficulty. Although the IMMPF idea is
not new [22-25], the IMMPF algorithm with quantized
measurements is firstly derived. Independent of the prior mode
information, both the state and the probability distribution of
the mode are estimated. The main contribution of this paper
includes: (i) overcoming the exponential computing difficulty
of the optimal MMSE approach and performing much more
efficiently and stable than MHMC method by cutting down the
computing expense and using only one tuning parameter (the
particle number) and (ii) better estimate performance than the
standard IMM directly using quantized measurements by
representing the posterior probability density under quantized
observation with the particle filter.
The paper is organized as follows. Firstly, in section 2,
problem formulation is presented. Subsequently, in section 3,
both the optimal and the quantized IMMPF estimation
framework are explained explicitly. Next, in section 4, the
complete algorithm of the quantized IMMPF is presented. The
simulation results and comparison are given in section 5.
Lastly, in section 6, some conclusions are drawn.
Science and Technology, Beijing, CO 100084 China (email: nyj10@
mails.tsinghua.edu.cn; {weidong, jyd}@mail.tsinghua.edu.cn).
State Estimation for Markov Jump Linear Systems with Quantized
Measurements: A Quantized IMMPF Algorithm *
Yingjun Niu, Wei Dong, Member, IEEE and Yindong Ji, Member, IEEE
2014 11th IEEE International
Conference on Control & Automation (ICCA)
June 18-20, 2014. Taichung, Taiwan
978-1-4799-2837-8/14/$31.00 ©2014 IEEE 941