sensor it is sampled by. Thus, the method studied in this paper will
be evolved by time-triggered mechanism, and updated by event-
triggered mechanism. The fusion estimate should be evolved in
each estimate evolution period and updated once a measurement
received by the fusion centre in the estimate evolution period. In
the estimate evolution period (i − 1)s, is , let N
i
be the amount of
the measurements received by the fusion centre and y
α
j
i
(β
j
i
)
be the
jth received measurement sampled by sensor be α
j
i
(1 ≤ α
j
i
≤ p) at
sampling time β
j
i
((i − 1)s < β
j
i
≤ is ).
For definiteness and without loss of generality, in the estimate
evolution period (k − 1)s, ks , the problem of sequential fusion
filtering under investigation is obtaining an estimate of z(ks) by the
filtering process of y
α
1
k
(β
1
k
)
and updating the estimate by the
filtering process of y
α
j
k
(β
j
k
), j = 2, …, N
k
. Therefore, the problem
of sequential fusion
ℋ
∞
filtering of the lth received measurement
in the estimate evolution period (k − 1)s, ks can be presented as
follows:
Problem 1: For a given scalar γ > 0 and the measurements
received by the fusion centre,
y
α
j
i
(β
j
i
)
if i = 1, …, k − 1 , j = 1, …, N
i
if i = k, j = 1, …, l
,
the goal is to obtain an estimate
z
^
l
(ks |ks)
of z(ks), which is
obtained or updated by filtering the lth received measurement in
the estimate evolution period (k − 1)s, ks , such that the following
sequential fusion
ℋ
∞
filtering performance index is satisfied: (see
(4)) where
e
z, j
(is) = z
^
j
(is| is) − z(is)
in which
z
^
j
(is| is)
is an
estimate of
z(is), and P
0
> 0 is a given symmetric weighting
matrix.
Remark 2: In the denominator of above performance index, the
process noise quadratic terms increase depending on the estimate
evolution period, while the measurement noise quadratic terms
increases when a measurement is received by fusion centre. As a
result, the proposed fusion method should be evolved by a time-
triggered mechanism and updated by a event-triggered mechanism.
Therefore, the provided performance index can be applied to the
multi-sensor systems, where sensors run with different operating
frequency. On the other hand, the real-time performance of the
fusion filtering can be ensured, without waiting for other
measurement sampled in the same estimate evolution period. For
single sensor time-invariant systems, such as the system (1)–(3)
with p
= 1, n
i
= 1, s = 1, the standard ℋ
∞
performance index is
usually give by Li and Sun [37]:
∑
k = 0
∞
∥ e
z
(k) ∥
2
< γ
2
(∑
k = 0
∞
∥ v(k) ∥
2
+ ∑
k = 0
∞
∥ w(k) ∥
2
)
. For the
single sensor time varying system, the standard indefinite quadratic
form [26, 32] is often presented as the form
∑
k = 0
N
∥ e
z
(k) ∥
2
< γ
2
∑
k = 0
N
∥ v(k) ∥
2
+ x
0
T
P
0
−1
x
0
+ ∑
k = 0
N
∥ w(k) ∥
2
.
For multi-sensor systems, such as the system (1)–(3) with
p
> 1, n
i
= 1, s = 1, the above two standard performance indexes
were developed as the form
∑
k = 0
∞
∥ e
z
(k) ∥
2
< γ
2
(∑
k = 0
∞
∥ V(k) ∥
2
+ ∑
k = 0
∞
∥ w(k) ∥
2
)
when
system parameters are constant [23, 38], and as the form
(1/
p)∑
k = 0
N
∥ e
z
(k) ∥
2
< γ
2
(∑
k = 0
N
∥ V(k ) ∥
2
+ x
0
T
P
0
−1
x
0
+ ∑
k = 0
N
∥ w(k) ∥
2
)
when system
parameters are time varying [39], where
V(k) = v
1
T
(k) v
2
T
(k) ⋯ v
p
T
(k)
T
.
They are presented for the fusion problem that there are fixed
number measurements received by fusion centre in each fusion
estimate evolving period and the received measurements are dealt
with in the parallel fusion frame. For multi-rate multi-sensor
systems, the number of the received measurements in each fusion
estimate evolving period is inconstant, as shown in Fig. 1 and
Table 1. Therefore, in the performance index (4), the unfixed N
i
is
introduced to denote the amount of the measurements received by
fusion centre in the fusion estimate evolving period (i − 1)s, is .
What's more, in (4),
l is utilised to denote the sequence number of
the current measurement received by fusion centre, which also
means the filtering for the current received measurement is the lth
update of the fusion estimate, in the sequential fusion frame.
3 Sequential finite horizon ℋ
∞
fusion filter
3.1 Analysis of the performance index
Consider the following indefinite quadratic form:
J
l
(ks) =
∑
τ = 1
ks
w
T
(τ, τ − 1)w(τ, τ − 1) + x
0
T
P
0
−1
x
0
+
∑
i = 1
k − 1
∑
j = 1
N
i
v
α
j
i
T
(β
j
i
)v
α
j
i
(β
j
i
) − γ
−2
e
z, j
T
(is)e
z, j
(is)
+
∑
j = 1
l
v
α
j
k
T
(β
j
k
)v
α
j
k
(β
j
k
) − γ
−2
e
z, j
T
(ks)e
z, j
(ks) .
(5)
The performance index (4) is satisfied if and only if
J
l
(ks) > 0 for all x
0
, w ∈ l
2
, v ∈ l
2
.
(6)
The indefinite quadratic inequality (6) is met, if and only if, the
indefinite quadratic form J
l
(ks)
has a positive minimum at a
stationary point.
For ease of description, we define the following notations:
F(i, j) = F(i, i − 1) × ⋯ × F(j + 1, j), i > j, and F(i, i) = I,
(see equation below)
Ψ
y, α
j
i
(β
j
i
) = H
α
j
i
(β
j
i
)F(β
j
i
, 0), Ω
y, α
j
i
k
(β
j
i
) = H
α
j
i
(β
j
i
)Γ
k
(β
j
i
),
Ψ
z, j
(is) = L(is)F(is, 0), Ω
z, j
k
(is) = L(is)Γ
k
(is),
(see equation below) According to the state-space model (1)–(3),
the expressions of
y
α
j
i
(β
j
i
)
and
z
^
j
(is| is)
can be written as
(see (7))
(see (8)) The augmented measurement
Y
l
k
is modelled as
sup
x
0
, w ∈ l
2
, v ∈ l
2
∑
i = 1
N
i
≠ 0
k − 1
∑
j = 1
N
i
e
z, j
T
(is)e
z, j
(is) + ∑
j = 1
l
e
z, j
T
(ks)e
z, j
(ks)
∑
i = 1
N
i
≠ 0
k − 1
∑
j = 1
N
i
v
α
j
i
T
(β
j
i
)v
α
j
i
(β
j
i
) + ∑
j = 1
l
v
α
j
k
T
(β
j
k
)v
α
j
k
(β
j
k
) + ∑
τ = 1
ms
w
T
(τ, τ − 1)w(τ, τ − 1) + x
0
T
P
0
−1
x
0
< γ
2
(4)
W
k
= w
T
(1, 0), w
T
(2, 1), …, w
T
(ks, ks − 1)
T
, Γ
k
(t) = F(t, 1), F(t, 2), …, F(t, t), 0, …, 0
ks − t
,
IET Control Theory Appl., 2017, Vol. 11 Iss. 3, pp. 369-381
© The Institution of Engineering and Technology 2016
371