136 C. Zhang et al. / Applied Mathematics and Computation 287–288 (2016) 134–148
The H
∞
adaptive smoothing problem for systems (1) can be stated as: given a scalar γ > 0, an integer l > 0, and the
observation { y (i ) }
k
i =0
, we find the adaptive smoothing estimate of z ( k ), denoted by
˘
z
(k
l
| k )(k
l
= k − l) , such that
sup
(x
0
,w, v ,θ) =0
N
k = l
˘
z
(k
l
| k ) − z(k
l
)
2
x
0
−
˘
x
0
2
P
−1
0
+
N
k =0
w (k )
2
+
N
k =0
v (k )
2
+
N
k =0
θ (k )
2
< γ
2
(4)
where x
0
is unknown and P
0
is a given positive definite matrix which reflects the relative uncertainty of the initial state
about the estimate
˘
x
0
= 0 .
Remark 1. In order to attenuate the disadvantage effect of the unknown time-varying parameter θ ( k ) to the H
∞
adaptive
smoother, it is reasonable that the denominator of the left-hand side of (4) contains the term of
N
k =0
θ (k )
2
.
Remark 2. Since the observation y ( k ) given in (1) only contains the information of { θ (i ) }
k −1
i =0
, it means that the filtering
algorithm cannot achieve the unknown parameter estimation, and the 1-step fixed-lag smoothing is the fastest estimate
algorithm. Therefore, we adopt the H
∞
fixed-lag smoothing algorithm to realize the hybrid estimation of state and unknown
parameter in this paper.
3. Main results
In this section, the H
∞
adaptive smoothing problem is reformulated as a positive minimum problem of an indefinite
quadratic form, and the minimum problem is discussed in Krein space by using innovation analysis approach. Subsequently,
the recursive H
∞
adaptive smoother is proposed in terms of two Riccati difference equations.
3.1. Solution of the minimum problem of indefinite quadratic form
Note that (x
0
, w, v , θ ) = 0 , it is obvious that (4) comes into existence if and only if the following inequality holds [28]
J
N
= x
0
2
P
−1
0
+
N
k =0
w (k )
2
+
N
k =0
v (k )
2
+
N
k =0
θ (k )
2
− γ
−2
N
k = l
v
z
(k
l
)
2
> 0 , ∀ (x
0
, w, v , θ ) = 0 (5)
where v
z
(k
l
) =
˘
z
(k
l
| k ) − z(k
l
) .
Moreover, it follows from (2) and (3) that
f (x (k ) , u (k )) θ(k ) − f (
ˆ
x (k ) , u (k )) θ(k )
≤ f (x (k ) , u (k )) − f (
ˆ
x (k ) , u (k )) θ(k )
≤ γ
2
γ
3
F (x (k ) −
ˆ
x (k )) (6)
Define
w
φ
(k ) = φ(x (k ) , u (k )) − φ(
˘
x
(k | k ) , u (k ))
w
f
(k ) = f (x (k ) , u (k )) θ (k ) − f (
˘
x
(k | k ) , u (k )) θ(k )
v
φ
(k ) = x (k ) −
˘
z
φ
(k | k ) ,
˘
z
φ
(k | k ) =
˘
x
(k | k )
v
f
(k ) = F x (k ) −
˘
z
f
(k | k ) ,
˘
z
f
(k | k ) = F
˘
x
(k | k ) (7)
where
˘
x
(k | k ) denotes the estimate of x ( k ) onto the linear space spanned by { y (i ) }
k
i =0
.
Remark 3. Note that
˘
x
(k | k ) is the estimate of x ( k ) onto the linear space spanned by { y (i ) }
k
i =0
. Therefore, φ(
˘
x
(k | k ) , u (k )) and
f (
˘
x
(k | k ) , u (k )) are obtainable on-line at time step k .
By considering (2), (6) and (7) , we introduce a new indefinite quadratic form
J
∗
N
= J
N
+
N
k =0
w
φ
(k )
2
− γ
2
1
N
k =0
v
φ
(k )
2
+
N
k =0
w
f
(k )
2
− γ
2
2
γ
2
3
N
k =0
v
f
(k )
2
(8)
Then, it follows from (2), (6) and (7) that J
∗
N
≤ J
N
. It is evident that if J
∗
N
> 0 then the H
∞
adaptive smoothing
problem (4) is satisfied, i.e., J
N
> 0. Therefore, in this paper, we will first find the minimum of J
∗
N
, and then choose
{
˘
z
φ
(k | k ) ,
˘
z
f
(k | k ) }
N
k =0
; {
˘
z
(k
l
| k ) }
N
k = l
such that the value of J
∗
N
is positive at its minimum. Specially, the minimum problem
of indefinite quadratic form J
∗
N
can be solved by employing a Krein space approach [28] .