q ¼
q
1
q
24
q
1
¼ cos ðh=2Þ q
24
¼
q
2
q
3
q
4
2
6
4
3
7
5
¼
^
n sin ðh=2Þ
ð1Þ
where
^
n is a unit vector corresponding to the axis of
rotation and h is the angle of rotation. The quaternion
kinematic equations of motion are derived by using the
spacecraft’s angular velocity x ¼½x
1
; x
2
; x
3
T
, given by
_
q ¼
1
2
XðxÞq ¼
1
2
EðqÞx ð2Þ
where XðxÞ, EðqÞ are defined by
XðxÞ¼
0 x
T
x ½x
; EðqÞ¼
q
T
24
q
1
I
33
þ½q
24
;
½x ¼
0 x
3
x
2
x
3
0 x
1
x
2
x
1
0
2
6
4
3
7
5
ð3Þ
Since a three degree-of-freedom attitude system is repre-
sented by a four-dimensional vector, the quaternions can-
not be independent. This condition leads to the following
normalization constraint as
q
T
q ¼ q
2
1
þ q
2
24
q
24
¼ 1 ð4Þ
In this paper, the attitude dynamics model is employed
to replace the inertial gyro model to describe the angular
velocity, which inevitably introduces uncertain model
errors. In addition, the perturbations form the inside of
satellite and outside from the space are unable to establish
or estimate accurately. For simplicity, the various kinds of
perturbations are included as the total model errors, which
is considered as d in the attitude dynamics model of this
paper, as
_
x ¼ J
1
½N x ðJxÞ þ J
1
d ð5Þ
where N denotes the total torque vector; J denotes the
moment of inertia tensor of the satellite, which needs to
be estimated in this paper; d represents the model errors
of the system and perturbations from the space. Therefore,
the attitude dynamics model without the gyro can be
expressed by
_
x ¼
_
q
_
x
¼
1
2
XðxÞq ¼
1
2
EðqÞx
J
1
½N x ðJxÞ þ J
1
d
"#
ð6Þ
The measurement data are obtained from the star sen -
sor, which is assumed to be of the form given by
y
j
¼ h
j
ðxÞ¼C
b
i
ðqÞr
j
þ m
j
; j ¼ 1; 2 ...n ð7Þ
where C
b
i
is the direction cosine matrix parameterized in
terms of the unknown true quaternion q; y
j
is the measure-
ment vector; r
j
is the reference vector of the star sensor in
the inertial frame; m
j
is the zero mean Gaussian white-noise
with covariance matrix r
2
I
33
; i is the corresponding num-
ber for difference reference vectors. Two reference vectors
of the star sensor are considered in this paper. Thus, the
measurement equations for the attitude estimate system
can be expressed as
y ¼
C
b
i
ðqÞr
1
C
b
i
ðqÞr
2
"#
þ
m
1
m
2
¼ hðxÞþv ð8Þ
where v is the zeros mean Gaussian white-noise process
with covariance R ¼ r
2
I
66
.
3. Double gain Kalman filter
The proposed double gain Kalman filter (DGKF) is
derived based on the Kalman filter theory by making full
use of the previous time-step’s measurement residuals. In
this section, it will be shown that the novel DGKF is
optimal in a mean-square error (MSE) sense and owns
some obvious advantages than the classical Kalman filter
(KF) by theoretical demonstrations.
Before the introduction of DGKF, some basic
parameters need to be defined firstly. The a priori state esti-
mate
^
x
k
is the state estimate at time t
k
just before employ-
ing the measurement y
k
in the estate update algorithm. The
a posteriori state estimate
^
x
þ
k
is the state estimate at time t
k
just after the state estimate update. The MSE performance
index is given as
J
k
¼ E ð
~
x
þ
k
Þ
T
~
x
þ
k
hi
ð9Þ
the optimal state solver of the above ind ex function is the
theoretical basis for the Kalman filter. Then, the a priori
and a posteriori estimation errors are defined, respectively
~
x
k
¼ x
k
^
x
k
;
~
x
þ
k
¼ x
k
^
x
þ
k
ð10Þ
Associated with the above estimation errors, we can define
the priori and posteriori state error covariance matrices
P
k
¼ E
~
x
k
ð
~
x
k
Þ
T
; P
þ
k
¼ E
~
x
þ
k
ð
~
x
þ
k
Þ
T
hi
ð11Þ
before and after the measurement update, respectively.
With the above preparation, the DGKF is derived based
on the discrete-time form of Eqs. (6) and (8) as
x
k
¼ f ðx
k1
Þþw
k1
y
k
¼ hðx
k
Þþv
k
ð12Þ
where k 2 N
0
is the discrete time. w
k1
and v
k
are indepen-
dent process and measurement Gaussian noises with zero
means and covariance Q
k1
and R
k
.
In order to derive the DGKF, the basic definitions of
residuals e
k
and e
þ
k1
are given respectively. The first resid-
ual e
k
is given between the actual measurement and predic-
tive measurement as
e
k
¼ y
k
^
y
k
¼ H
k
ðx
k
^
x
k
Þþv
k
¼ H
k
~
x
k
þ v
k
;
H
k
¼
@h
@x
x¼
^
x
k
ð13Þ
L. Cao et al. / Advances in Space Research 60 (2017) 499–512 501