Académique Documents
Professionnel Documents
Culture Documents
Shovan Bhaumik
Samar Bhattacharya
I NTRODUCTION
459
P ROBLEM F ORMULATION
A. Process Model
We assume the target is falling vertically downwards on
the earth as shown in Figure 1. Further, We assume (i)
drag and gravity are acting in straight line, (ii) lift force is
negligible small, (iii) earth is at and stationary, and (iv)
altitude independent due to gravity. Based on above mentioned
assumptions, the target follows kinematic equations as [1], [2],
[4].
h = v
(1)
v =
(h)gv 2
+g
2
(2)
= 0,
(3)
2
2016 IEEE First International Conference on Control, Measurement and Instrumentation (CMI)
0].
III.
altitude, where, a1 = 1.754, and a2 = 1.49 104 . Now, the
process model of missile dynamics have been discretized its
state xk is shown as
xk+1 = f (xk ) + wk ,
(4)
3
2
0
q3 3 q3 2
2
(6)
Qk =
,
q3
q3
0
2
0
0
q4
where, the tuning parameters q3 in (m2 /s3 ) and q4 in
(kg 2 m2 s5 ) are to be selected by the designers.
In Figure 2 to Figure 4, plotted the typical missile trajectory
of position, velocity and acceleration, where no process noise
is taken such that q3 = q4 = 0. Initial height, velocity
and ballistic coefcient are taken as 60960m, 3048m/s and
460
(7)
N ON LINEAR F ILTERS
(9)
Yk|k1 = Hxjk|k
(10)
(11)
where
Lk|k1 =
1
1
2
Zk|k1 ) (Zk|k1
Zk|k1 )
[(Zk|k1
N 1
N
... (Zk|k1
Zk|k1 )]
2016 IEEE First International Conference on Control, Measurement and Instrumentation (CMI)
and
Zk|k1
C. Gauss-Hermite Filter
N
1
j
=
Z
N j=1 k|k1
(12)
Kk = Pk|k1 GT (GPk|k1 GT + Rk )1
(13)
I1p ].
(14)
j
are calculated using measurement noise distribuwhere, yk,0
j
tion, yk,0 (yk , Rk ).
(15)
n
2n
1
Ai f ( 2i [ui ]) ,
I(f )
2n(n/2) i=1
(16)
i =1
n
d
= (1)n e n +n e = 0
d
(17)
n !( + n + 1)
and Ai =
is the Gauss-Laguerre weights.
2
i [L
n (i )]
where = (n/2 1).
Now, from
equation (16), we get a cubature quadrature
points j = 2i [ui ] and its corresponding weights wj as,
wj =
(19)
1
n !( + n + 1)
Ai =
2
2n(n/2)
i [L
n (i )]
Calculate mean
f (qi )wi
i=1
(18)
IN
...
i1 =1
(21)
in =1
In n-dimensional integral system with m-point single dimensional Gauss-Hermite quadrature rule has total mn multidimensional points. Hence the lter suffers from the curse of
dimensionality problem.
D. Sparse grid Gauss-Hermite Filter
The SGHF [13], [14] uses sparse-grid quadrature rule for
nding points and weights. It works on deterministic sampling
point approach. For univariate, GHF and SGHF have same
quadrature points and weights. In multivariate, GHF faces the
curse of dimensionality problem. SGHF uses Smolyaks rule
for generation of points and weights, which is free from the
curse of dimensionality problem. The sparse-grid integral is
dened as,
f (x)(x; 0, In )dx In,L (f )
(22)
Rn
In,L (f ) =
L1
q=Ln
L1q
(1)L1q Cn1
Nqn
(23)
where In,L is the approximation of the n-dimensional integral
of function f (.) with Gaussian weights (x, 0, In ) with accuracy level L N. It gives the exact solution for all polynomial
upto degree (2L 1). C denotes as a binomial coefcient,
dened by Ckn = n!/k!(n k)!, where n! denote as a factorial
of n. denotes the tensor product. Iij is a single dimensional
quadrature rule where ij is the accuracy level. It means
that Iij gives exact result upto polynomial of degree 2ij 1.
The accuracy level sequence
n
{ : j=1 ij = n + q} f or q 0
n
(24)
Nq =
f or q < 0
2016 IEEE First International Conference on Control, Measurement and Instrumentation (CMI)
L1
....
qsn Xin
L1q
L1q
Cn1
n
(25)
w sj }
j=1
Xn,L =
where
(26)
Figure 5: Mean position error for various lters.
q=Ln Nqn
S IMULATION R ESULTS
A. Initialisation
In this paper, we have simulated with initial position
h = 60960m, velocity v = 3048m/s and ballistic coefcient, where beta distribution is used for initializing the
ballistic coefcient () with both parameters are 1.1, i.e.
Beta(1.1, 1.1) and, lower and upper limit boundaries
are 10000kg/ms2 and 63000kg/ms2 respectively. We have
taken the initialization from [1],[2], where ballistic coefcient
is distributed randomly because we have no prior information
about the targets shape, size and mass.
The initialization of the lter with its mean
T
x
0|0
= [60960 3048 mean(0 )] , and covariance
Rk
0
Rk
R
R
P0|0 = k 2 k 0 , where mean 0 and standard
2
0
0
2
deviation are generated randomly from beta distribution.
B. Simulation results
The above mentioned missile tracking problem has been
solved with EnKF with ensemble size 50, UKF, CKF, CQKF-2,
GHF-3 and SGHF-2. In this problem, we have taken sampling
time = 0.1sec and simulation is observed upto 30sec. The
tuning parameter of process noise for estimator and truth are
given in [1] with q3 = q4 = 5. The measurement noise
covariance(Rk ) is assumed as Rk = 2002 . Above discussed
lters are simulated upto 100 Monte Carlo runs.
The mean error of estimator are plotted for position in
Figure 5, velocity in Figure 6 and ballistic coefcient in Figure
7 for 100 MC runs respectively. The standard deviation of
estimator are plotted for position in Figure 8, velocity in
Figure 9 and ballistic coefcient in Figure 10 for 100 MC
runs respectively. The performance of these lters have been
compared on the basis of estimation accuracy, computational
time and missed distance. GHF-3 has smallest position error
shown in Figure 8. In Figure 9, velocity error has spike occured
between 5s and 25s. Largest spike has occured in SGHF-2
at 16s. In Figure 10, GHF-3 has lowest ballistic coefcient
error. Missed distance is the last point error calculation from
462
C ONCLUSION
2016 IEEE First International Conference on Control, Measurement and Instrumentation (CMI)
6
7
12
27
7
50
Normalized
computational time
1
1.054979
1.619338
2.999348
1.097639
6.299187
Missed distance
of position error (m)
32.18156
29.98065
30.59586
26.90438
29.93088
29.19445
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
463