Spacecraft Attitude & Rate Estimation by an
Adaptive Unscented Kalman Filter
Manasi Das, Smita Sadhu, Tapan Kumar Ghoshal
Dept. of Electrical Engg.
Jadavpur University
Kolkata – 700032, India.
Abstract—This paper presents a scheme for spacecraft
attitude and rate estimation based on an Adaptive Unscented
Kalman Filter (AUKF). The integrated attitude determination
system here consists of rate gyros and attitude sensors as the
measurement units. The process equations are formed combining
the information from the rotational dynamics and kinematics of
the spacecraft. As the measurement noise statistics of the rate
gyros and the attitude sensors are often unknown, so an adaptive
filter is employed here, which determines the measurement noise
covariance (R) adaptively. Further, the choice of UKF facilitates
with its high accuracy and derivative-free implementation. Two
versions of the AUKF algorithm are shown here depending on the
residual sequences and the innovation sequences. The residual
based proposed AUKF comprises of two phase of measurement
update. Further, a provision to regulate the speed of adaptation
through a modulating factor, is also incorporated in the presented
adaptive schemes. Positive definiteness and the diagonal structure
of the adapted measurement noise covariance (R) is also ensured
here. Simulation results show the superiority of the AUKF based
attitude estimation scheme compared to the nonadaptive filter
based schemes in terms of the estimation errors.
Keywords—Adaptive filter; Integrated attitude determination
system; Spacecraft rotational dynamics; Unscented Kalman filter
I. INTRODUCTION
This paper has presented a spacecraft attitude and rate
estimation technique based on a novel Adaptive Unscented
Kalman Filter (AUKF). Unfortunately, the most widely used
Extended Kalman Filtering (EKF) scheme for nonlinear system
is based on a sub-optimal implementation [1]. This may
sometime degrade the estimation performance or even may lead
to divergence. Whereas, Unscented Kalman Filter (UKF) use a
carefully selected set of sample points known as sigma points
to map the probability distribution, providing better accuracy
and faster convergence [2] compared to the EKF especially in
strongly nonlinear systems with large initial errors. Calculation
burden of the UKF is also less as it doesn’t require complex
Jacobian calculations like the EKF.
On the other hand, nonadaptive filters may it be EKF or
UKF require a priori knowledge of process and measurement
noise covariances. If, the prior knowledge of the noise
covariances used in the filter model does not correspond to the
actual one, estimation accuracy may be seriously affected. To
obviate this difficulty here in our work adaptive filtering
framework is utilized.
The idea of employing AUKF for state estimation is
relatively new which is found in few recent works. Most of the
existing AUKF algorithms have utilized scaling factors [3]-[9]
to make the UKF adaptive. In [3] an AUKF is employed to
estimate the attitudes of a pico satellite, where the process
noise covariance (Q) is adapted by a scaling factor only when
faulty condition is detected depending on the residual
sequences. In [4] the same authors have used another AUKF
for the same purpose, where instead of the process noise
covariance (Q), the measurement noise covariance (R) is
scaled by a scaling matrix, at the presence of measurement
malfunction. But, the positive definiteness is not guaranteed in
this proposed scaling method. To ensure the positive
definiteness another type of R scaling method is shown in [5].
But this scaling method fails to adapt R when the filter R is
initialized with a higher value than the true one. In [6], both Q
and R scaling, as presented in [3] and [5] respectively, are
executed depending on the type of the fault (whether process
fault or measurement fault). Another Q and R scaling methods
are shown in [7], where the scaling factors are determined by
an optimal Downhill Simplex search technique. In [8] another
type of AUKF is involved, where the scaling of predicted
error covariance (P
-
) and the Kalman gain (K) is implemented
by utilizing two different scaling factors. This method also
fails to adapt when the filter R is initialized with a higher
value than the true one. In [9] only the predicted error
covariance (P
-
) is scaled to make the UKF adaptive. Both [8]
and [9] are unable to estimate the noise covariances and rely
on the prior knowledge about them. AUKF based on complex
MIT adaptation rule may be found in [10]. But the
implementation of this method is often unrealistic due to its
huge partial derivative calculations.
Among the all existing AUKF techniques the simplest
technique is to adapt the Q or R directly from the innovation
sequences or the residual sequences. In [11] an AUKF for
GPS based position estimation problem is presented where, Q
is estimated directly from the innovation sequences. In [12] a
residual based R adaptive AUKF algorithm is shown.
Although, the defining formula ((14) in [12]) rebels with the
corresponding R-adaptive Kalman filter relation. Another
innovation based AUKF algorithm with both Q and R
adaptation and a forgetting factor is shown in [13]. However
in this algorithm only the innovation sequence of the k
th
instant is utilized.
The first author acknowledges the financial support of Council of
Scientific & Industrial Research (CSIR), India.
2014 International Conference on Control, Instrumentation, Energy & Communication(CIEC)
978-1-4799-2044-0/14/$31.00©2014IEEE
96