DISTRIBUTED WEIGHTED FUSION ESTIMATORS 593
series. But it has weak reliability, i.e., the filter would fail if there were a faulty
sensor during the estimation process. Also, it does not consider the packet drop-
ping. So far, distributed fusion estimation is seldom reported for sensor networks
with random time delays and packet dropping.
In this paper, we study the distributed fusion estimation for discrete linear
stochastic systems in sensor networks. Each sensor estimates the system state
based on its own measurement data, and then sends its estimate to the fusion
centre for fusion, however, with random delays and packet dropping. Based on the
optimal fusion algorithm weighted by matrices in the linear minimum variance
sense [14], we give two optimal distributed weighted fusion estimators, which
involve the weighted fusion of different step predictions from different sensors.
The cross-covariance matrix of different step prediction errors between any two
local estimates is derived. Local estimators with lower precision due to the large
time delays or consecutive packet dropping are not fused in the fusion centre to
reduce the computational burden because they do not bring more improvement
in precision to the fusion estimator. We can implement this strategy by setting a
precision requirement, then determining the gate threshold for time delays or the
number of consecutive packet dropouts to satisfy the given precision constraint
for all local estimators. That means that only the local estimators with time delays
or a number of consecutive packet droppings less than the determined gate values
are fused. We obtain a suboptimal distributed weighted fusion estimator that has
approximate precision compared with the optimal distributed weighted fusion
estimator and has a reduced computation cost.
The rest of this paper is organized as follows. The problem formulation is given
in Section 2. The optimal distributed fusion estimators are given and the cross-
covariance matrix between any two local estimates is derived in Section 3. In
Section 4, the gate threshold for time delay or the number of consecutive packet
droppings for some given precision requirement is determined, and a suboptimal
fusion strategy is introduced. In Section 5, simulation results for a tracking system
with four sensors are reported. The conclusions are drawn in Section 6.
2. Problem formulation
Consider the discrete time-invariant linear stochastic control system with L sen-
sors
x(t + 1) = x(t) + w(t ) (1)
y
(i)
(t) = H
(i)
x(t) + v
(i)
(t), i = 1, 2,...,L, (2)
where x(t) ∈ R
n
is the state, y
(i)
(t) ∈ R
m
i
is the measurement, w(t) ∈ R
r
and
v
(i)
(t) ∈ R
m
i
are white noises, and , , H
(i)
are constant matrices with com-
patible dimensions. Superscript (i) denotes the ith sensor, and L is the number of
sensors.