Anda di halaman 1dari 13

Aerodynamic parameter estimation using

adaptive unscented Kalman lter


M. Majeed and Indra Narayan Kar
Department of Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, India
Abstract
Purpose The purpose of this paper is to estimate aerodynamic parameters accurately from ight data in the presence of unknown noise characteristics.
Design/methodology/approach The introduced adaptive lter scheme is composed of two parallel UKFs. At every time-step, the master UKF
estimates the states and parameters using the noise covariance obtained by the slave UKF, while the slave UKF estimates the noise covariance using the
innovations generated by the master UKF. This real time innovation-based adaptive unscented Kalman lter (UKF) is used to estimate aerodynamic
parameters of aircraft in uncertain environment where noise characteristics are drastically changing.
Findings The investigations are initially made on simulated ight data with moderate to high level of process noise and it is shown that all the
aerodynamic parameter estimates are accurate. Results are analyzed based on Monte Carlo simulation with 4000 realizations. The efcacy of adaptive
UKF in comparison with the other standard Kalman lters on the estimation of accurate ight stability and control derivatives from ight test data
in the presence of noise, are also evaluated. It is found that adaptive UKF successfully attains better aerodynamic parameter estimation under the same
condition of process noise intensity changes.
Research limitations/implications The presence of process noise complicates parameter estimation severely. Since the non-measurable process
noise makes the system stochastic, consequently, it requires a suitable state estimator to propagate the states for online estimation of aircraft
aerodynamic parameters from ight data.
Originality/value This is the rst paper highlighting the process noise intensity change on real time estimation of ight stability and control
parameters using adaptive unscented Kalman lter.
Keywords Aircraft, Aerodynamics, Noise, Stability (control theory), Adaptive lter, Parameter estimation, Unscented Kalman lter,
Aerodynamic parameters
Paper type Research paper
Nomenclature
C
D
; C
L
; C
m
non-dimensional derivatives
F
e
thrust (N)
I
y
inertia about y-axis (kg m
2
)
m mass of the aircraft (kg)
P state error covariance
q pitch rate (radian/s)
Q state noise covariance
R measurement noise covariance
S surface area of aircraft (m
2
)
u input vector
V true airspeed (m/s
2
)
w state noise
x state vector
z observation vector
c wing chord (m)
q dynamic pressure (pascal)
Greek symbols
a angle of attack (radian)
b aerodynamic parameters
u pitch angle (radian)
Q unknown parameters
a
1
; b
1
; k scaling parameters of UKF
s
T
engine inclination (radian)
n additive measurement noise
h diagonal elements of noise covariance
Subscripts
0 initial value
m measured quantity
k index
Superscripts
a augmented state vector
c computation of covariance
m computation of mean
Introduction
Recent advances in computational power have allowed the use
of online aerodynamic parameter estimation techniques in
varied applications such as recongurable or adaptive ight
control, and system health monitoring. Online modelling is a
key technology for autonomous controller to maintain ight
stability and high performance in uncertain environment and
in the presence of process and measurement noise.
Robust and adaptive control methods suffer from issues
related to the real time convergence, and the complications
involved in their real-time implementation (Wittenmark and
Astrom, 1984). These problems necessitate the development
The current issue and full text archive of this journal is available at
www.emeraldinsight.com/1748-8842.htm
Aircraft Engineering and Aerospace Technology: An International Journal
85/4 (2013) 267279
q Emerald Group Publishing Limited [ISSN 1748-8842]
[DOI 10.1108/AEAT-Mar-2011-0038]
267
of a new online estimation algorithm for the purpose of
controller design that addresses the situation more directly.
To this end, autonomous control methods on the basis of
model-reference have become the focus of research, and basic
technology and online modelling method has attracted more
and more research attention.
Neural networks (NN) and NN-based self-learning were
proposed as the most effective approaches for the online
modeling of an unmanned vehicle as it does not require initial
values (Pesonenet al., 2004; Pedro and Kantue, 2011). However,
the problems involved in NN, such as training data selection,
online convergence, robustness, reliability, and real time
implementation, limit its application in real systems. In recent
years, sequential estimation has become an important approach
for online modelling and model-reference control with
encouraging achievements (Haykin and deFreitas, 2004).
Online or recursive system identication techniques handle
ight data as it is measured through onboard sensors and
estimate the required aerodynamic derivatives in real-time.
Measured ight data can contain considerable amount of
noise, furthermore there might be biases and unobserved
states in the system model which must be estimated; hence
ltering techniques are generally employed. The most popular
nonlinear ltering technique is the extended Kalman lter
(EKF) (Grewal and Andrews, 1993). Although widely used,
EKFs have some deciencies, including the requirement of
differentiability of the state dynamics as well as susceptibility
to bias, hard to tune and implement when dealing with
signicant nonlinearities and exhibits divergence in estimates.
On the contrary, unscented Kalman lter (UKF) uses the
nonlinear model directly instead of linearizing it (Julier and
Uhlmann, 2004).
The UKF was developed with the underlying assumption that
approximating a Gaussian distribution is easier than
approximating a nonlinear transformation (Julier and Uhlmann,
2004). The UKF uses deterministic sampling to approximate
the state distribution as a Gaussian randomvariable (GRV). The
UKF has the same level of computational complexity as that of
EKF, both of which are within the order O (L3) (Julier and
Uhlmann, 2004). Since the nonlinear models are used without
linearization, the UKF does not need to calculate Jacobians or
Hessians, and can achieve second-order accuracy, whereas the
accuracy of the EKF is of the rst order (Xiong et al., 2006).
However, the performance of UKF closely depends on the
prior knowledge of noise covariance. Recent study shows the
use of UKFhas possible advantages for aerodynamic parameter
estimation from ight data (Chowdhary and Jategaonkar,
2009). The UKF can only achieve a good performance under
the assumption that such prior knowledge of noise meets the
situation well (Julier and Uhlmann, 1995). But in practice, the
assumptions are usually not totally satised, and
the performance of the UKF might be seriously downgraded
from the theoretical performance or could even diverge. To
avoid these problems, an adaptive lter may be applied, which
automatically tunes the lter parameter to adapt insufciently
known a priori lter statistics. There have been many
investigations in the area of adaptive lter and parameter
estimation. Maybeck (1979) used a maximum-likelihood
estimator for designing an adaptive lter that could estimate
the system-error covariance matrix. Katebi et al. (1985)
introduced state space LQG adaptive autopilot for ship
control. Lee and Alfriend (2004) modied the Maybecks
methods by introducing a window-scale factor. A multiple
Kalman lter based adaptive stochastic ltering method was
applied for the online aircraft ight path reconstruction with
estimation of noise statistics (Celso-Braga et al., 2007). The
estimator was introduced in UKF to estimate its noise statistics
through maximizing the posteriori density function (Zhao and
Wang, 2009), and modied Sage estimator was also utilized
to estimate the system noise variance adaptively (Shi et al.,
2009). Song and Han (2008) were designed a MITrule based
adaptive UKFalgorithmto update the covariance of the process
uncertainties online by minimizing the cost function. But it
introduces a relative large computational burden. Recently,
Kalman lter based adaptive UKF algorithm is introduced for
the actuator failure estimationof Rotorcraft UAV(Juntonget al.,
2012). Moreover, Chanying and Guo (2010) have established
a new critical theorem for global stabilization of adaptive
nonlinear system. The new automated adaptive algorithms are
integrated into the UKF and can be applied to the nonlinear
system.
Nonlinear dynamics of aerospace vehicles and the presence
of considerable noise and biases in measurements demand that
a nonlinear ltering algorithm be used (Jategaonkar and
Plaetschke, 1989). Traditionally, the EKF has been used
for parameter estimation purposes. In this paper, an online
innovation-based adaptive UKFis usedto estimate aerodynamic
parameters from ight data. The used adaptive lter composed
of two parallel UKFs, called as master UKF and slave UKF.
The master UKF estimates both the states and aerodynamic
parameters of aircraft while the slave one estimates the diagonal
elements of the noise covariance matrix for the master UKF.
This approach of noise covariance identication based adaptive
UKF is previously applied in mobile robot systems (Song et al.,
2007). By estimating the noise covariance, this adaptive method
is able to compensate the estimation errors resulting from the
insufcient knowledge of the noise statistics.
Performance evaluation of adaptive UKF was done with use
of both simulated and ight test data. The simulated data
were generated with moderate to high level of process noise.
A 4,000 samples Monte Carlo simulation was done and the
results were very encouraging to apply on ight test data.
The main contribution of this paper is the application of
adaptive lter method to estimate aerodynamic parameters
recursively from ight data, and successfully reduces the
dependency of the estimation performance on the accurate
knowledge of the system noise statistics. The adaptive UKF
was compared with ltering algorithms of the EKF (Grewal
and Andrews, 1993), standard UKF (the simplied version of
the UKF), augmented UKF (the complete version of the
UKF) to estimate aerodynamic derivatives from ight test
data. It was shown that adaptive UKF is superior to classical
UKF in terms of fast convergence and estimation accuracy
under the variation of noise characteristics.
The paper is organized as follows. A brief review of the
recursive parameter estimation (RPE) is presented in second
section. The applied adaptive UKF structure is illustrated in
third section. The ight simulated and ight test data have
applied to the adaptive lter and other RPE methods in fourth
section. The estimation results are analyzed in fth section
and concluding remarks are presented in the last section.
Recursive parameter estimation
The basic ideas of practical engineering applications of RPE
originate from control applications in chemical and thermal
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
268
power-generating industrial process, where online adaptation
of model parameters is desired to increase the overall plant
efciency. In such applications, we encounter not only changing
operating conditions but also several different types of system
disturbances. In the case of aircraft, the primary motivation for
recursive estimation could be to obtain immediate knowledge
about anaircraft model, whichis essential for designing adaptive
control. These real time applications necessarily call for
recursive/online estimation methods.
The system dynamics are represented in generic
continuous-discrete state space form in equation (1)
(Jategaonkar, 2006):
_ xt f xt; ut; b Fwt; xt
0
x
0
yt hxt; ut; b
zk yk Gnk; k 1; . . . ; N
8
>
>
<
>
>
:
1
where x is state vector with initial value x
0
at time t
0
, u is the
input vector, y is the observation vector, f and h are the general
nonlinear functions, z is the measurement vector sampled at N
discrete time steps with a xed sampling time Dt and k is
discrete time index. The measurement noise vector n is assumed
to be sequence of independent zero mean white Gaussian noise.
The matrices F and G represent the distribution matrices for
additive state and measurement noise, respectively. The
unknown parameter vector Q consists of system parameters b,
the measurement biases Dz, and the trim estimates formulated
as input biases Du and is represented in equation (2):
Q
T
b
T
; Dz
T
; Du
T
2
Estimation of these parameters through the ltering approach is
an indirect procedure, consisting of transforming the parameter
estimation problem into a state estimation problem. This is done
byaugmentingthesystemstatevector byintentionallydeningthe
unknownparameters as additional state variables. So the constant
systemparameters Qconsideredas output of anauxiliary dynamic
system presented in equation (3):
_
Q 0 3
The augmented state vector is then dened in equation (4) as:
x
a

x
Q
" #
4
where the augmented variables are denoted by subscript a. The
extended system is represented as:
_ x
a
f
a
x
a
t; ut F
a
W
a
t

f xt; ut; b
0
" #

F 0
0 0
" #
wt
0
" #
yt h
a
x
a
t; ut
zk yk Gvk
8
>
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
>
:
5
The value of F and G, i.e. the process and measurement noise
distributionmatrices must be specieda priori. The measurement
noise matrix G can be calibrated using laboratory measurements
of sensors to ensure good noise ltering. The process noise matrix
F is however more difcult to determine and a trial and error
method is normally employed If no other suitable method is
found. Since the inaccurate assumptions of F will lead to poor
performance or even divergence of the lter (Julier and Uhlmann,
1995), it is necessary to identify the process noise covariance by
using adaptive algorithm, which attempts to adapt the unknown
noise characteristics (Julier and Uhlmann, 2004).
The adaptive UKF structure
The applied adaptive scheme is composed of two parallel
UKFs as shown in Figure 1. At every time-step, the master UKF
estimates the states and parameters using the noise covariance
obtained by the slave UKF, while the slave UKF estimates the
noise covariance using the innovations generated by the master
UKF. The master UKF can also work independently without
the slave one. Thus, dual-UKFstructure is reducedtoa standard
UKF with x noise covariance. The setting of master UKF does
not need any updates while activating/deactivating the slave one,
which indicates that the slave UKF can be shut down to reduce
the computational burden when the system statistics do not
change a lot.
The master UKF
In the adaptive scheme, the calculation of the master UKF is
the same as that of a classical UKF. To illustrate the UKF
procedure, we consider a discrete time representation widely
used in the applications of the UKF. The discrete time
nonlinear state space model is given by:
x
k1
f
d
x
k
; Q
k
; u
k
w
k
z
k
y
k
v
k
(
6
y
k
h
d
x
k;
Q
k
; u
k
7
where x
k
is the (n
k
1) state vector, Q
k
the (n
q
1) vector of
unknownparameters, u
k
the (n
u
1) vector of exogenous inputs,
y
k
the (n
y
1) model output vector at time k,f
d
and h
d
the
corresponding state and output functions. w
k
and n
k
are,
respectively, the disturbance and sensor noise vector, which are
assumed to be zero mean Gaussian white noise with their
covariance matrices QandR. Themeasurement vector is denoted
by z
k
. The augmented state vector of the size (n
a
1) is given by:
x
a
k

x
T
k
Q
T
k
h i
T
8
where the superscript a denotes the augmented state vector,
and n
a
n
x
n
q
is the total number of states:
Figure 1 The adaptive UKF structure
Master UKF
Slave UKF
Time update
Measurement
update
Measurement
update
z
k
Measurement
v
k
Innovation
State
Noise
Covariance
Time update
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
269
Master UKF: initialization
^ x
a
0
E x
a
0

E ^ x
T
0
Q
T
0

T
^ x
T
0
Q
T
0

T
^
P
a
0
E x
a
0
2 ^ x
a
0
x
a
0
2 ^ x
a
0

T
n o

P
x0
0
0 P
Q0
" #
8
>
>
>
>
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
>
>
>
>
:
9
Master UKF: sigma points calculation and time update
~ x
a
k

^ x
a
k
; ^ x
a
k
2

n
a
l
^
P
a
k
;
q
^ x
a
k

n
a
l
^
P
a
k
q
~ x
a
k1
^ x
a
k

R
tk1
tk
f x
a
t; ukdt
~ x
k1

P
2na
i0
W
m
i
~ x
a
i;k1
~
P
k1

P
2na
i0
W
c
i
~ x
a
i;k1
2 ~ x
k1
h i
~ x
a
i;k1
2 ~ x
k1
h i
T
Q
x
a
8
>
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
>
:
10
where:
W
m
0

l
nal
W
c
0

l
nal
1 2a
2
1
b
1

W
m
i
W
c
i

1
2nal
i 1; . . . ; 2n
a
l a
2
1
n
a
k 2n
a
8
>
>
>
>
>
>
<
>
>
>
>
>
>
:
11
Master UKF: measurement update
Y
k1
h
d
~ x
a
k1
; u
k1

~ y
k1

X
2na
i0
W
m
i
Y
i;k1
P
~ y~ yk1

X
2na
i0
W
c
i
Y
i;k1
2 ~ y
k1
Y
i;k1
2 ~ y
k1

T
R
x
a
P
~ x~ yk1

X
2na
i0
W
c
i
~ x
a
i;k1
2 ~ x
k1
h i
Y
i;k1
2 ~ y
k1

T
K
K1
P
~ x~ yk1
P
21
~ y~ yk1
m
k1
z
k1
2 ~ y
k1
^ x
k1
~ x
k1
K
k1
m
k1
^
P
k1

~
P
k1
2K
k1
P
~ y~ yk1
K
T
k1
12
The variables in equations (9)(12) are dened as followings.
W
i
is a set of scalar weights, a
1
determines the spread of the
sigma points around the estimated x
a
and is usually set as
0.0001 # a
1
# 1 whereas k is secondary scaling parameter
usually set to zero for state or (3 2n
a
) for parameter
estimation. The constant b
1
is used to incorporate part of the
prior knowledge of the statistics of x, while b
1
2 is optimal
for Gaussian distributions. The process and measurement
noise covariance are represented by Q
x
a
and R
x
a
, respectively,
whose diagonal elements are required to be estimated by the
slave UKF.
The slave UKF
There are six parameters in UKF, which are the initial state
^ x
0
, initial covariance
^
P
0
, process-noise covariance Q,
measurement-noise covariance R, and unscented transform
(UT) parameters a
1
and b
1
. The inuence of the initial state
and covariance will become asymptotically negligible as more
and more data are processed. The selection for a
1
and b
1
has
little impact on improving the estimate accuracy of the UKF if
the order of the system under consideration is not high. As a
priori knowledge, initial covariance P
0
, the covariance
matrices Q and R are most important to the performance
and stability of the UKF. The P
0
can be chosen based on the
physical knowledge of variation in state variables, but Q and R
are associated with uncertainties in the environment. In
principle, an adaptive lter can estimate both R and Q, and
their theoretical justication were well described by Myers
and Tapley (1976). However, adaptive ltering algorithms
that try to update both the observational noise and the system
noise are not robust, since it is not easy to distinguish between
errors in R and Q (Blanchet et al., 1977). The measurement-
noise statistics are relatively well known compared to the
system-model error. Therefore, selecting noise covariance
matrix Q is most important to maintain the performance and
stability of the UKF.
In this section, the slave UKF is described to estimate the
noise covariance matrix Q. The criterion for estimation of Q is
to minimize the difference between the lter-computed and
the actual innovation covariance. From equation (12) of the
master UKF, the computed innovation covariance can be
obtained as:
~
S
k
P
~ y~ yk
diag
X
2na
i0
W
c
i
Y
i;k
2 ~ y
k
Y
i;k
2 ~ y
k

T
R
x
a
" #
13
The time-averaged approximation of actual innovation
covariance is dened as:
S
k

1
N
X
k21
ik2N
m
i
m
T
i
14
where N is the size of the estimation window, m
i
is the
innovation and can be written as:
m
i
z
i
2 ~ y
i
15
where z
i
and ~ y
i
are, respectively, the real measurement and its
estimated value. Thus, the condition for the estimated Q is
the minimization of S
k
2
~
S
k

, i.e. S
k
2
~
S
k
< 0:
S
k
E m
k
m
T
k

P
~ y~ yk
16
From equation (12), K
k
P
~ y~ yk
K
T
k

~
P
k
2
^
P
k
can be written as:
P
~ y~ yk
K
T
k
K
k

21
K
T
k

~
P
k
2
^
P
k
K
k
K
T
k
K
k

21
17
Referring to the master UKF equations, we can get
~
P
k
;
^
P
k
; K
k
as follows:
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
270
P
~ y~ yk

K
T
k
K
k

21
K
T
k
X
2na
i0
W
c
i
~ x
a
i;k
2~ x
k
h i
~ x
a
i;k
2~ x
k
h i
T
Q2
^
P
k
!
K
k
K
T
k
K
k

21


K
k
X
2na
i0
W
c
i
~ x
a
i;k
2~ x
k
h i
~ x
a
i;k
2~ x
k
h i
T
2
^
P
k
!

K
T
k


K
k
Q

K
T
k
18
where

K
k
K
T
k
K
k

21
K
T
k
so that

K
k

21
K
k
. From
equations (16) and (18), innovation sequence can be written as:
E{m
k
m
T
k
} P
~ y~ yk


K
k
X
2na
i0
W
c
i
~ x
a
i;k
2 ~ x
k
h i
~ x
a
i;k
2 ~ x
k
h i
T
2
^
P
k
!

K
T
k


K
k
Q

K
T
k
19
Thus, the noise covariance matrix Q is estimated from equation
(19) as:
^
Q K
k
E m
k
m
T
k

K
T
k
2
X
2na
i0
W
c
i
~ x
a
i;k
2 ~ x
k
h i
~ x
a
i;k
2 ~ x
k
h i
T
2
^
P
k
! 20
Usually, the process-noise covariance Q is a diagonal matrix.
Therefore, the estimation of Qcan be simplied as the estimation
of its diagonal elements. Thus, we use the h
i
k
to denote the
diagonal elements of matrix Q
x
k
, i.e.:
Q
x
k

h
n1
k
0
.
.
.
.
.
.
.
.
.
0 h
nx
k
2
6
6
6
4
3
7
7
7
5
and Q
x
a
k

Q
x
k
0
0 0
" #
21
If the dynamics of h is known, state equation of the slave UKF is:
h
k1
f h
k
w
hk
22
If the dynamics of h is unknown, it can be modelled as a non-
correlated random drift vector:
h
k1
h
k
w
hk
23
where w
hk
is the Gaussian white noise with zero mean. The
innovationcovariancegeneratedbythemaster UKFis takenas the
observation signal for the slave UKF, and then according to
equation (12), the observation model can be described as:
~
S
k
gh
k

diag
X
2na
i0
W
c
i
Y
i;k
2 ~ y
k1
Y
i;k
2 ~ y
k

T
R
x
a
" # 24
The measurement of
~
S
k
received by the slave UKF is:
S
k
diag m
k
m
T
k

; m
k
z
k
2 ~ y
k
Therefore, the recursive algorithm of the slave UKF can be
formulated as:
Slave UKF: initialization
^ h
0
Eh
0

^
P
h0
Eh
0
2 ^ h
0
h
0
2 ^ h
0

8
<
:
25
Slave UKF: sigma points calculation and time update
q
k

^ h
k
; ^ h
k
2

n l
^
P
hk
;
q
^ h
k

n l
^
P
hk
q
~
q
K1
q
k
~ h
k1

X
2n
i0
W
m
hi
~
q
i;K1
~
P
hk1

X
2n
i0
W
c
hi

~
q
i;K1
2 ~ h
k1

~
q
i;K1
2 ~ h
k1

T
Q
h
26
where:
W
m
h0

l
nl
W
c
h0

l
nl
1 2a
2
1
b
1

W
m
hi
W
c
hi

1
2nl
i 1; . . . ; 2n
l n a
2
1
21

8
>
>
>
>
>
>
<
>
>
>
>
>
>
:
Slave UKF: measurement update
6
k1
g
d

~
q
k1

~
S
k1

P
2n
i0
W
m
hi
6
i;k1
P ~
S
~
Sk1

P
2n
i0
W
c
hi
6
i;k1
2
~
S
k1
6
i;k1
2
~
S
k1

T
R
h
P
~ h
~
Sk1

P
2n
i0
W
c
hi

~
q
i;k1
2 ~ h
k1
6
i;k1
2
~
S
k1

T
K
hk1
P
~ h
~
Sk1
P
21
~
S
~
Sk1
^ h
k1
~ h
k1
K
hk1
S
k1
2
~
S
k1

^
P
hk1

~
P
hk1
2K
hk1
P ~
S
~
Sk1
K
T
hk1
8
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
:
27
where n is state dimension. Q
h
and R
h
are the process and
measurement noise covariance, respectively.
Aerodynamic parameter estimation from
simulated ight data
To evaluate the efcacy of the adaptive UKF algorithm,
aircraft responses pertaining to longitudinal motion are
generated through the simulation of aircraft dynamics
models. An independent process and measurement noise
vectors are generated using pseudo-random noise generators
(Jategaonkar and Plaetschke, 1989). The state noise matrix is
assumed to be diagonal and an appreciable level of process
noise is incorporated in the system dynamics. Note that state
variables are only affected by the state noise, and the control
input is noise free. A total of 30s of data with a sampling time
0.025 s are generated.
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
271
Aircraft dynamic model for parameter estimation
The longitudinal motion of aircraft responses are represented
using the following postulated aircraft models (Klein and
Moreli, 2006):
_ a 2
qS
mV
C
L
q
g
V
cosa 2u 2
Fe
mV
sina s
T

_
u q
_ q
qS c
Iy
C
m

Fe
Iy
l
tx
sins
T
l
tz
coss
T

8
>
>
>
<
>
>
>
:
28
where the lift, and pitching moment coefcients (C
L
; C
m
) are
modelled as:
C
L
C
L0
C
La
a C
Lq
q c
2V
C
Lde
d
e
C
m
C
m0
C
ma
a C
mq
q c
2V
C
mde
d
e
8
<
:
29
Here the state variables are angle of attack a, pitch angle u,
and pitch rate q. The elevator deection d
e
and thrust F
e
are
considered to be control input to aircraft system dynamics
equation (28). The mass of aircraft is m and its geometrical
parameters consist of wing area S, wing chord c, moment of
inertia I
y
, and the inclination angle of the engines s
T
. The
dynamic pressure q is given by q 0:5rV
2
, where r is the
density of air and V is the true airspeed. The estimation
algorithm has used equations (28) and (29) as state equations,
and corresponding observation equations are given in
equation (30):
a
m
a
u
m
u
q
m
q
_ q
m
2
qS c
I
y
C
m

F
e
I
y
l
tx
sins
T
l
tz
coss
T

a
zm
2
qS
m
C
L
cosa 2
F
e
mV
sins
T
30
The unknown parameter vector Q need to be estimated is
consisting of the aerodynamic derivatives, which is given by:
Q
C
L0
C
La
C
Lq
C
m0
C
ma
C
mq
C
mde
h i
T
Results and analysis for simulated data
In the absence of ight test data, generated ight simulated
data with moderate to high level of process noise to estimate
the aerodynamic parameters using adaptive UKF algorithm.
The simulated data contain adequate level of process noise
similar to aircraft ying in turbulence (Iliff, 1978). Figure 2
shows the ight simulated and model estimated responses.
This time history plots for the model output from Adaptive
UKF show good agreement except at the rst data point for
the estimated signals az. This is because, the initial state
error covariance matrix (P
o
) specied may be too large for the
system states or parameters. The sigma points are computed
based on initially specied state and state error covariance
matrix. For the subsequent data points the sigma points are
generated using the updated covariance, which are computed
by the adaptive UKF algorithm, and are more realistic. This
problem is specic to UKF only.
Figure 3 compares the performance of the adaptive UKF
algorithmover the simple UKFfor the purpose of aerodynamic
parameter estimation. There is initial mismatch in marginal
level of their estimates due to the use of very low initial values
of Q in adaptive lter. But time progresses, the accurate
agreement is observed between adaptive UKF and classical
UKF estimates of the parameters. It shows that adaptive UKF
does not match correctly with all classical UKF estimates,
namely C
La
and C
mde
as shown in Figure 3, and there is an error
between them, which is within the limits of considering the
noise variation. It is worth to note that the classical UKF uses
accurate value of Q, but adaptive UKF estimates accurately
compared to the classical UKF even without knowledge of
Q. The performance of the adaptive lter was further analyzed
based on a Monte Carlo simulation with 4,000 realizations.
In the state estimation model, diagonal elements of noise matrix
are selected randomly for the 4,000 realizations. Figure 4 shows
the consistency of aerodynamic parameter estimation from
simulated data using adaptive UKF for the varied process noise
properties. Mean value of all estimated parameters for 4,000
iterations are listed in Table I. The good agreement between all
estimated aerodynamic parameters and their true value is very
encouraging to apply on ight test data.
Aerodynamic parameter estimation from ight
data
Open accessible ight test data of the research aircraft HFB-
320 (Figure 5) were used to estimate aircraft lift, drag, and
pitching moment coefcients (Jategaonkar, 2006) and
demonstrated the efcacy of the adaptive UKF algorithm
for aerodynamic parameter estimation from ight data. For
the use of ight test data, ight tests were carried out to excite
longitudinal motion of research aircraft through a multi-step
elevator input resulting in short period motion and a pulse
input leading to phugoid motion (Jategaonkar, 2006). These
recorded ight test data has been used without any smoothing
or ltering; the angle of attack (a) was calibrated using ight
path reconstruction techniques (Majeed and Kar, 2010). The
thrust was computed prior to parameter estimation and used
in the estimation process as an input variable.
Aircraft dynamic model for parameter estimation
The longitudinal model of the research aircraft is represented
in equation (31) and contains one additional state variable of
true airspeed V to the model equation (28) used for
generating the ight simulated data in the previous section.
This state variable V leads to additional nonlinearities due to
its inversion in the state equation of a, and increases the
number of parameters by incorporating drag derivatives as
given in the following postulated aircraft state models
(Jategaonkar, 2006; Chowdhary and Jategaonkar, 2009):
_
V 2
qS
m
C
D
g sina 2u
Fe
m
cosa s
T

_ a 2
qS
mV
C
L
q
g
V
cosa 2u 2
Fe
mV
sina s
T

_
u q
_ q
qS c
Iy
C
m

Fe
Iy
l
tx
sins
T
l
tz
coss
T

8
>
>
>
>
>
>
<
>
>
>
>
>
>
:
31
where the lift, drag, and pitching moment coefcients
(C
L
; C
D
; C
m
) are modelled as:
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
272
Figure 3 Performance of adaptive and standard UKF algorithms
0.22
0.2
12
11
10
0.229
0.2285
0.228
22.8
23
5.65
5.6
0.6
0.5
0.4
1.155
1.16
1.185
1.382
1.384
0 10 20 30 40 0 10 40
0 10 20 30 40 0 10 20 30 40
0 10 20
Time in sec
Time in sec
30 40 0 10 20 30 40
0 10 20 30 40 0 10 20 30 40
Adap UKF
True Value
UKF
C
L
0
C
L
L
a
C
L
L
a
C
m
m

e
C
m
m

e
C
L
q
C
m
0
C
m
q
Figure 2 Histories of output variables for the generic transport aircraft
20
0
20
50
0
40
20
0
20
40
50
0
50
40
20
0
20
40
10
5
d
e
l
e
(

)
a
z
(
m
/
s
2
)
q
p
(

/
s
2
)
q
(

/
s
)

)
0
0 5 10 15 20 25 30
50
Time in sec
Simulated, and
estimated
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
273
C
D
C
D0
C
DV
V
Vo
C
Da
a
C
L
C
L0
C
LV
V
Vo
C
La
a
C
m
C
m0
C
mV
V
Vo
C
ma
a C
mq
q c
2Vo
C
mde
d
e
8
>
>
>
<
>
>
>
:
32
where V
0
is initial value of V. The dynamic pressure q
0:5rV
2
contains state variable V which introduces
nonlinearity in estimation process as it multiplies with all
derivatives needs to be estimated. The following measurement
model is used to estimate the aerodynamic derivatives
including drag coefcients:
V
m
V
a
m
a
u
m
u
q
m
q
_ q
m
2
qS c
I
y
C
m

F
e
I
y
l
tx
sins
T
l
tz
coss
T

a
xm

qS
m
C
X

F
e
m
coss
T
a
zm

qS
m
C
Z

F
e
m
sins
T
33
Figure 5 The DLR HFB-320 research aircraft
Figure 4 Performance of adaptive UKF algorithm in Monte Carlo simulation
0.215
0.21
0.205
5.62
5.61
5.6
11
10.5
10
0.48
0.47
0.46
0.2283
0.2282
1.161
1.1612
1.1614
22.98
22.99
23
1.3825
1.383
1.3835
0 500 1,000 1,500 2,000
Number of iterations
2,500 3,000 3,500 4,000
C
m

e
C
L

e
C
L

C
L
q
C
m

C
m
0
C
L
0
C
m
q
Table I Lift force and pitching moment derivatives
M 5 0.25, H 5 5,000 ft
Parameters True value Mean estimated values
C
L0
0.2102 0.2095 (7.14 10
204
)
*
C
La
5.6275 5.6165 (0.0017)
C
Lq
10.5249 10.5584 (0.1097)
C
Lde
0.4671 0.4712 (0.0076)
C
m0
0.22826 0.2283 (1.9 10
204
)
C
ma
21.1613 21.1613 (4.603 10
204
)
C
mq
222.9920 222.9920 (0.0289)
C
mde
21.3831 21.3830 (0.002)
Note:
*
Mean values of standard deviation of estimates
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
274
where the longitudinal and vertical force coefcients C
X
and
C
Z
are given by:
C
X
C
L
sina 2C
D
cosa
C
Z
2C
L
cosa 2C
D
sina
(
34
Thus, the unknown parameter vector Q need to be
estimated is consisting of the aerodynamic derivatives and is
given by:
Q
C
D0
C
DV
C
Da
C
L0
C
LV
C
La
C
m0
C
mV
C
ma
C
mq
C
mde
h i
T
Results and analysis for ight test data
The adaptive UKF algorithm was applied to estimate the
aerodynamic parameters from the ight data. We have
introduced initial state x
0
(106.02, 0.11, 0.15, 20.003) for
master UKF and x
20
(1.2856, 5.86 10
2005
, 2.01 10
2006
,
3.02 10
2005
) for slave UKF. Similarly, state propagation error
covariance matrixP
0
diag(10, 10, 10, 10) for master UKFand
P
20
diag (1, 1, 1, 1) for slave UKF. Figure 6 shows the ight
measured and model estimated responses for the HFB-320
researchaircraft. Themodel estimatedsignals axandazshow
large deviations fromthe rst fewdata points. It may be due to the
initial state error covariance matrix(P
0
) speciedmaybetoolarge
for the system states or parameters. Similar type of initial
mismatch was shown in Figure 2 for the case of simulated data.
Apart from this, the agreement between ight measured and
model estimated responses is good.
The similar parameter estimation procedure was carried
out using EKF, UKF, and UKFaug algorithms and compared
their performance for the purpose of aerodynamic parameter
estimation (Chowdhary and Jategaonkar, 2009). Here, we
repeated the applications of these EKF, UKF, and UKFaug
algorithms for the purpose of comparison with Adaptive UKF
algorithmand showing the agreement in estimated parameters.
The details of the estimation methods (EKF, and UKFaug) are
not described here for brevity, and they are illustrated in
(Chowdhary and Jategaonkar, 2009).
Figure 7 shows the performance of these RPE methods for
the purpose of aerodynamic parameter estimation from ight
data. It shows that convergence of the estimates for the adaptive
UKF is similar to the usual UKF and other methods, so that
the use of adaptive UKF is appreciable without knowing Q.
Table II compares the numerical values in four decimal places of
the estimates of the parameters arrived at with the different
methods. It is clearlyseenthat the parameter estimates of all these
methods are in close vicinity of one another, and all numerical
values are ingoodagreement. The performance of these methods
Figure 6 Histories of output variables for HFB-320 aircraft
114
104
7.5
5
10
2
2.5
2.5
8
0
8
1.2
0.7
7.5
11.5
0
0
a
z
(
m
/
s
2
)
q
d
o
t
(

/
s
)
q
(

/
s
)

)
a
x
(
m
/
s
2
)
v
(
m
/
s
)
10 20 30 40 50 60
Time in sec
Measure and Estimated

e
(

)
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
275
and their close agreement canlargely be attributedto anaccurate
mathematical model of the system under consideration. The
performance analysis of EKF and UKF are discussed broadly in
the context of aerodynamic parameter estimation (Chowdhary
and Jategaonkar, 2009). Here, we are emphasizing the
application of adaptive UKF to obtain immediate knowledge
about an aircraft model under variation of process noise due to
the external disturbance. The next section illustrates the
credibility of adaptive lter.
The dynamic changes of process noise
The estimationaccuracyof the adaptive UKFwithrespect tothe
situation of changing process-noise covariance is tested. The
change of the true process-noise intensity is assumed as:
Q
To
diag{1:2856 5:861210
25
0:203510
25
3:019910
25
}; t ,20sec
Q
T
diag{1:285610
23
5:861210
28
0:203510
28
3:019910
28
}; t $20sec
8
>
>
>
>
>
<
>
>
>
>
>
:
35
Filter error method (FEM) represents the most general
stochastic approach to aircraft parameter estimation
(Jategaonkar and Plaetschke, 1989, 1990; Jategaonkar, 1993)
and estimates state noise distribution matrix F additional to the
system parameters to calculate process noise covariance
Q FF
T
Dt
21
, Dt is sampling time. In UKF, the prior
knowledge of the process noise covariance is selected as
Q Q
T0
, which is the true process noise covariance computed
Figure 7 Performances of RPE methods
Table II Comparison of parameter estimates
Parameters RPE methods
FEM Adapt. UKF UKF UK Faug EKF
Computational time 7.55 5.83 11 2.56
C
D0
0.1226 (2.45) 0.1226 (2.64) 0.1226 (2.64) 0.1239 (2.55) 0.1235 (2.50)
C
DV
20.0645 (3.95) 20.0642 (4.27) 20.0642 (4.27) 20.0653 (4.09) 20.0652 (4.01)
C
Da
0.3201 (2.26) 0.3197 (2.40) 0.3197 (2.40) 0.3160 (2.37) 0.3191 (2.33)
C
L0
20.0929 (21.10) 20.0871 (22.91) 20.0871 (22.91) 20.0990 (20.03) 20.0853 (23.49)
C
LV
0.1487 (11.10) 0.1468 (11.42) 0.1468 (11.42) 0.1569 (10.61) 0.1440 (11.69)
C
La
4.3278 (1.08) 4.2894 (1.14) 4.2894 (1.14) 4.3032 (1.13) 4.3027 (1.14)
C
m0
0.1119 (3.27) 0.1115 (4.29) 0.1115 (4.29) 0.1152 (3.40) 0.1115 (4.28)
C
mV
0.0040 (82.10) 0.0045 (92.19) 0.0045 (92.19) 0.0022 (153.88) 0.0046 (90.48)
C
ma
20.9679 (1.12) 20.9703 (1.54) 20.9703 (1.54) 20.9832 (1.24) 20.9712 (1.54)
C
mq
234.7098 (2.27) 235.3628 (2.82) 235.3628 (2.82) 235.0981 (2.27) 234.9372 (2.85)
C
mde
21.5291 (1.27) 21.5395 (1.65) 21.5395 (1.65) 21.5517 (1.31) 21.5328 (1.65)
Note: The values in parenthesis denote standard deviation values in percent
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
276
by ofine FEM(Jategaonkar and Plaetschke, 1990) and it is get
changed after 20 s, i.e. Q Q
T
. But in the case of adaptive
UKF, process-noise intensity xed to the very small quantity
for the credibility to select arbitrarily small value of Q:
Q
adp
To
diag{1:2856 10
214
5:8612 10
219
0:2035 10
219
3:0199 10
219
}; t , 20 sec
Q
adp
T
diag{1:2856 10
217
5:8612 10
221
0:2035 10
221
3:0199 10
221
}; t $ 20 sec
8
>
>
>
>
>
<
>
>
>
>
>
:
36
Thus, the Qof the adaptive lter is randomly selectedtovery small
value as Q Q
adp
T0
without a priori knowledge of noise statistics
and it is get changed to Q Q
adp
T
after the time of 20s. The state-
estimation errors of the classical UKF, and the adaptive UKF,
under the same condition of the process-noise intensity change,
are shown in Figure 8. It shows that, under incorrect noise
information for the case of Q change after 20s, the classical
UKF cannot produce optimal estimates as showing the case of
Q xed due to the violation of the optimality conditions.
On the contrary, the estimation errors in the adaptive UKF are
quickly overcome at the time of 20s and are almost the same as
their previous value for the cost of additional computation in
estimation of process noise (Figure 8(b)). The estimated Q value
in average for the last ten iterations of adaptive UKF is given by:
^
Q
adp
avg
diag{1:2856 5:8612 10
25
2:0035
10
26
3:0199 10
25
}
37
The aerodynamic pitching moment derivatives estimates by
standard UKF, and the adaptive UKF under the same condition
of the process-noise intensity change, are shown in
Figures 9 and 10, respectively. We can see that, by the classical
UKF, the aerodynamic parameter estimations happen to
divergence after the time of 20s due to the priori covariance
setting of UKF fails to meet the true values. By the application of
adaptive UKF, on the other hand, the estimated parameters
Figure 8 State-estimation errors with the time varying process noise
(a) Standard UKF
(b) Adaptive UKF
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
277
converge to same as before (Figure 10) even though the process
noise covariance was changed.
In real application, there is a possibility of the error in the
assumption of the prior knowledge of noise (say order of 10
23
).
This leads tothe xing of very low/highvalue of Qwithreference
to its true value, and degrades the performance of the UKF
throughparameter divergence as showninFigure 9. It is a major
issue in online ight stability and control parameter estimation,
which has addressed by the adaptive UKF. It is ensure that by
online estimating the process noise covariance, the adaptive
UKF successfully reject the inuence caused by the
incorrect priori covariance of process noise and achieve better
estimates of aerodynamic parameters under varying noise
statistics.
Conclusions
This paper described the application of adaptive UKF
algorithm to estimate aircraft aerodynamic parameters from
noisy ight data. Initially, it was successfully applied to the ight
simulated data with moderate to high level of process noise and
the results were analyzed based on a Monte Carlo simulation
with 4,000 realizations. Subsequently, the performance of the
adaptive UKF was demonstrated to identify the aircraft system
dynamics from ight test data, which was superior to the
classical UKF in terms of fast convergence and estimation
accuracy while changing the noise statistics. Therefore, the
potential use of adaptive UKF is favoured to the design of
adaptive ight controller for the recursive estimation of
unobserved states in the aircraft system model and unknown
aerodynamic parameters. It is worth to note that the two UKFs
are independent in the adaptive UKFstructure. Thus, the slave
UKF can be replaced by another simple lter such as Kalman
lter to save the computational burden further. This may be
considered as a future work, if the efcacy of proposed adaptive
lter is signicant for the cost of additional computation in
estimation of process noise.
Figure 9 UKF-divergence of parameter estimates due to changing process noise
0
0.95
1
1.05
25 30 35 40 45 50 55 60
60 50 40 30
Time in sec
20 10 0
1
2
3
20
25
30
35
C
m
q
C
m

UKF Q change
UKF +
UKF
FEM
Figure 10 Adaptive UKF-convergence of parameter estimates while changing process noise
2
1
0
1
2
3
20
25
30
35
40
0.9
1
1.1
0 10 20 30
Time in sec
40 50 60
25 30 35 40 45 50 55 60
C
m
q
C
m

AdpUKE
AdpUKF+
AdpUKF
FEM
AdpUKE
AdpUKF+
AdpUKF
FEM
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
278
References
Blanchet, I., Frankignoul, C. and Cane, M. (1977),
A comparison of adaptive Kalman lters for a tropical,
Pacic Ocean Model, Vol. 125 No. 1, pp. 40-58.
Celso-Braga, D.M., Hemerly, E.M. and Carlos, S.L. (2007),
Adaptive stochastic ltering for online aircraft ight path
reconstruction, Journal of Aircraft, Vol. 44 No. 5,
pp. 1546-1558.
Chanying, L. and Guo, L. (2010), A new critical theorem for
adaptive nonlinear stabilization, Automatica, Vol. 46,
pp. 999-1007.
Chowdhary, G. and Jategaonkar, R.V. (2009), Aerodynamic
parameter estimation from ight data applying extended
and unscented Kalman lter, Aerospace Science and
Technology, Vol. 14 No. 2, pp. 106-117.
Grewal, M.S. and Andrews, A.P. (1993), Kalman Filtering
Theory and Practice, Prentice-Hall, Upper Saddle River, NJ.
Haykin, S. and deFreitas, N. (2004), Special issue on
sequential state estimation, Proceedings of the IEEE.
Iliff, K.W. (1978), Identication and stochastic control of an
aircraft ying in turbulence, Journal of Guidance, Control,
and Dynamics, Vol. 1 No. 2, pp. 689-695.
Jategaonkar, R.V. (1993), A comparison of output error and
lter error methods from aircraft parameter estimation
results, Proceedings of the NAL-DLR Symposium on System
Identication, DLR-Mitt, 93-1463-87.
Jategaonkar, R.V. (2006), Flight Vehicle System Identication:
ATime Domain Methodology, AIAA Progress in Astronautics
and Aeronautics Series, AIAA, Reston, VA.
Jategaonkar, R.V. and Plaetschke, E. (1989), Algorithm for
aircraft parameter estimation accounting for process and
measurement noise, Journal of Aircraft, Vol. 26 No. 4,
pp. 360-372.
Jategaonkar, R.V. and Plaetschke, E. (1990), Identication
of moderately nonliear ight mechanics systems with aditive
process and measurement noise, Journal of Guidance,
Control and Dynamics, Vol. 13 No. 2, pp. 277-285.
Julier, S. and Uhlmann, J.K. (2004), Unscented ltering and
nonlinear estimation, Proceedings of the IEEE.
Julier, S.J. and Uhlmann, J.K. (1995), A new approach for
ltering nonlinear system, paper presented at American
Control Conference.
Juntong, Q., Song, D., Wu, C., Han, J. and Wang, T. (2012),
KF-based adaptive UKF algorithm and its applications for
rotorcraft UAV actuator failure estimation, International
Journal of Advanced Robotic Systems, Vol. 9, pp. 132-141.
Katebi, M.R., Jgrimble, M. and Byrne, J. (1985), LQGadaptive
autopilot design, Proceedings of the 7th IFAC Symposium on
Identication and System Parameter Estimation, York.
Klein, V. and Moreli, E.A. (2006), Aircraft system
identication theory and practice, AIAA, Education
Series, AIAA, Reston, VA.
Lee, D.J. and Alfriend, K.T. (2004), Adaptive sigma point
ltering for state and parameter estimation, paper
presented at AIAA/AAS Astrodynamics Specialist
Conference and Exhibit, RI.
Majeed, M. andKar, I.N. (2010), Calibrationof air data systems
using extended Kalman ltering technique, International
Journal of Aerospace Innovations, Vol. 2 Nos 1/2, pp. 133-147.
Maybeck, P.S. (1979), Stochastic Models, Estimation, and
Control, Academic Press, New York, NY.
Myers, K.A. and Tapley, B.T. (1976), Adaptive sequential
estimation with unknown noise statistics, IEEETransactions on
Automatic Control, Vol. 21 No. 4, pp. 520-523.
Pedro, J.O. and Kantue, P. (2011), Online aerodynamic
parameter estimation of a miniature unmanned helicopter
using radial basis function neural networks, paper
presented at 8th Asian Control Conference, Kaohsiung,
Taiwan, May.
Pesonen, U., Steck, J. and Rokhsaz, K. (2004), Adaptive
neural network inverse controller for general aviation safety,
Journal of Guidance, Control, and Dynamics, Vol. 27 No. 3,
pp. 434-443.
Shi, Y., Han, C. and Liang, Y. (2009), Adaptive UKF for
target tracking with unknown process noise statistics,
paper presented at 12th International Conference on
Information Fusion, Seattle, WA, USA.
Song, Q. and Han, J.D. (2008), An adaptive UKF algorithm
for the state and parameter estimations of a mobile robot,
Acta Automatica Sinica, Vol. 34 No. 1.
Song, Q., Jiang, Z. and Jinda-Da, H. (2007), Noise covariance
identication based adaptive UKF with application to mobile
robot systems, paper presented at IEEE International
Conference on Robotics and Automation, Italy.
Wittenmark, B. and Astrom, K.J. (1984), Practical Issues in
the implimentation of self-tuning control, Automatica,
Vol. 20 No. 5, pp. 595-605.
Xiong, K., H, Y. and Zhang, C.W. (2006), Performance
evaluation of UKF-based nonlinear ltering, Automatica,
Vol. 42, pp. 261-270.
Zhao, L. and Wang, X. (2009), An adaptive UKF with noise
statistic estimator, paper presented at 4th IEEE Conference
on Industrial Electronics and Applications (ICIEA 2009),
Xian.
About the authors
M. Majeed completed his B.Tech
(Instrumentation & Control) from Calicut
University in 1998 and M.Tech (Control
Engineering) from IIT Delhi, India in 2002. He
is a research scholar at IIT Delhi and working as
senior scientist in CSIR-National Aerospace
Laboratories, Bangalore. His research interests
are aircraft parameter estimation, nonlinear systemidentication,
contraction based stability analysis, and data fusion. M. Majeed
is the corresponding author and can be contacted at:
majeed_md_123@rediffmail.com
Indra Narayan Kar completed his M.Tech and
PhD (Electrical) from IIT Kanpur in 1991 and
1997, respectively. He is working as Professor at
the Department of Electrical Engineering, Indian
Institute of Technology Delhi. His research
interests are robust and intelligent control,
nonlinear control, and system identication.
To purchase reprints of this article please e-mail: reprints@emeraldinsight.com
Or visit our web site for further details: www.emeraldinsight.com/reprints
Aerodynamic parameter estimation using adaptive UKF
M. Majeed and Indra Narayan Kar
Aircraft Engineering and Aerospace Technology: An International Journal
Volume 85 Number 4 2013 267279
279