Gaussian Sum Filter for State Estimation of Markov
Jump Nonlinear System
Li Wang, Yan Liang, Xiaoxu Wang and Linfeng Xu
School of Automation, Northwestern Polytechnical University
Xi’an, Shaanxi, China, 710072
Abstract—This paper proposes the Gaussian sum filter-
ing (GSF) framework for the state estimation of Markov
jump nonlinear systems (MJNLSs). Through present-
ing the Gaussian sum approximations about the model-
conditioned state posterior probability density function
(PDF) and the model-conditioned measurement posterior
predictive PDF, a general GSF framework in the mini-
mum mean square error (MMSE) sense is derived. The
Minor Gaussian-set design is utilized to merge the Gaus-
sian components at the beginning, which can effectively
limit the computational requirements. Simulation results
demonstrate that the proposed method performs almost
as well as the interacting multiple model particle filter
(IMM-PF) but with much lower computational cost.
Index Terms—Markov jump nonlinear systems; Gaus-
sian sum approximation; Moment matching; Polynomial
interpolation
I. INTRODUCTION
Markov jump systems (MJSs) involve both time-
evolving and event-driven mechanisms, which have been
extensively used to model the systems with variable
structures caused by sudden environment changes, sys-
tem noises and/or failures occurred in components etc.
in many fields such as target tracking [1, 2], seismic
signal processing [3], process monitoring and fault de-
tection [4]. For the state estimation problem of MJSs,
Bar-shalom has derived the recursive optimal estimator
for the general MJSs based on Bayesian theory [5].
However, for most nonlinear dynamical and measure-
ment models, an exact and recursive computation of the
state PDF is intractable. Consequently, one must rely
on suboptimal or approximate nonlinear filters. Much
attention has been paid on the state estimation problem
of nonlinear system. The filtering of MJSs can be divided
into the following two categories.
The first category is the state estimation of Markov
jump linear systems (MJLSs), in which the state space
Research supported in part by NSFC through grants 61135001,
61374023, 61074179, 61075029 and 61374159. Aerospace Fund
through grant 2014-HT-XGD, and NWPU Basic Science Fund
through grant 3102014KYJD030.
is linear and the system model is switched according
to a Markov chain. Since one has to match a filter
to each model, an exponentially increasing number of
filters are needed, which makes the optimal approach
impractical [2]. Hence, suboptimal techniques must be
utilized to avoid the exponentially increasing number
of histories. A simple-minded suboptimal technique is
to keep the
N histories with the largest probabilities
and normalize the probabilities. Another effective way is
merging technique used in GPB1 (generalized pseudo-
Bayesian of order 1), GPB2 and IMM [2, 5, 6], in which
the posterior PDFs can be represented exactly by a sum
of Gaussians. Learn from [5], the IMM estimator, which
(for
r models) consists of r model-matched sub-filters
for the model-based state estimation, has essentially
the same computational requirements as the GPB1 but
performs almost as well as the GPB2, which consists of
r
2
model-matched sub-filters.
The second category is the state estimation of Markov
jump nonlinear systems (MJNLSs), in which the state
space is nonlinear. For this case, such techniques might
work for mild nonlinearities by simply replacing the
Kalman filter (KF) with the extended or unscented
Kalman filter (EKF or UKF) in the IMM framework.
Thus, if the state PDF is inherently non-Gaussian or
the model exhibits any significant nonlinearities, the
suboptimal EKF and UKF can diverge or fail. In recent
years, the IMM-PF, which is a Particle filter (PF) for
stochastic hybrid system with a mixing step at the be-
ginning of each estimation cycle in the IMM framework,
has been studied in many literatures [7]-[12]. PF is a re-
cursive numerical implementation of the exact Bayesian
filtering scheme where the posterior distribution of the
state is represented by a set of random samples with
associated weights, which are propagated through the
dynamic system by using the importance sampling to
sequentially update the posterior distribution. However,
the PF is computation-intensive in real time applications
and may face the problem of particle impoverishment.
This problem is particularly compounded when particle
filters are used in hybrid system estimation.