You are on page 1of 9

Acta Astronautica 61 (2007) 831 – 839

www.elsevier.com/locate/actaastro

A constant gain Kalman filter approach for the prediction


of re-entry of risk objects
A.K. Anilkumara,∗ , M.R. Ananthasayanamb , P.V. Subba Raoc
a Vikram Sarabhai Space Centre, Indian Space Research Organization, Trivandrum 695 022, India
b Department of Aerospace Engineering, Indian Institute of Science, Bangalore 560 012, India
c Applied Mathematics Division, Vikram Sarabhai Space Centre, Indian Space Research Organization, Trivandrum 695 022, India

Received 28 June 2005; received in revised form 16 November 2006; accepted 25 January 2007
Available online 30 April 2007

Abstract
The accurate estimation of the predicted re-entry time of decaying space debris objects is very important for proper planning
of mitigation strategies and hazard assessment. This paper highlights the implementation strategies adopted for the online re-
entry prediction using Kalman filter approach with constant gains with the states being the semi-major axis, eccentricity and
ballistic coefficient and using the measurements of the apogee height and perigee height derived from the Two Line Elements
provided by agencies like USSPACECOM. Only a very simple model is utilised for the orbit propagation and a basic feature
of the present approach is that any unmodellable state and measurement errors can be accounted for by adjusting the Kalman
gains which are chosen based on a suitable cost function. In this paper we provide the details of validating this approach by
utilising three re-entries of debris objects, namely, US Sat. No. 25947, SROSS-C2 Satellite and COSMOS 1043 rocket body.
These three objects re-entered the Earth’s atmosphere on 4th March 2000, 12th July 2001 and 19th January 2002, respectively.
© 2007 Elsevier Ltd. All rights reserved.

1. Introduction on the TLE sets provided in [3] are the orbital


parameters together with rate of mean motion decay
The accurate estimation of the final orbital character- and an equivalent ballistic coefficient B ∗ . The physical
istics of decaying objects is of considerable importance parameters of the objects like mass, area of cross sec-
for predicting the re-entry time to plan mitigation strate- tion, shape and dimensions are not accurately available
gies and hazard assessment. Space debris studies [1,2] and further any modelling of the atmosphere in which
demand assessment of orbital life times of a large num- the objects decay is also uncertain and variable. Also,
ber of objects resulting from on-orbit fragmentation or due to the tumbling of the bodies and the complex
collision. The usual database available for the re-entry gas–surface molecular interaction the changing ballis-
time or the orbital lifetime prediction of large debris ob- tic coefficient B ∗ further complicates the prediction of
jects is the set of Two Line Elements (TLEs) provided by re-entry time.
the agencies like USSPACECOM [3]. The information The orbital propagator utilised in this study is a very
simple model [4,5], as provided in Appendix A, which
∗ Corresponding author. Fax: +91 471 270 4134. accounts for only the atmospheric drag effect on the
E-mail addresses: anilk_ak@yahoo.com (A.K. Anilkumar), orbital decay of the object. This model estimates only
sayanam@aero.iisc.ernet.in (M.R. Ananthasayanam), the semi-major axis and eccentricity decay in one revo-
pulipakasrao@yahoo.co.in (P.V. Subba Rao). lution by considering only constant scale height during
0094-5765/$ - see front matter © 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.actaastro.2007.01.063
832 A.K. Anilkumar et al. / Acta Astronautica 61 (2007) 831 – 839

Nomenclature v measurement noise


w process noise
a semi-major axis in km
X state vector
A/m area to mass ratio
Z measurement vector
B ballistic coefficient (=CD × A/m)
B∗ drag term in TLE data, related to B Greek letters
CD non-dimensional drag coefficient  innovation
e eccentricity  density
ha apogee height  state transition matrix
hp perigee height
H measurement matrix Superscripts
J normalised innovation cost function ∼ value before update
K Kalman gain matrix ˆ value after update
N number of measurements T transpose of a matrix
P0 initial state covariance −1 inverse of a matrix
Q process noise covariance
Subscripts
R measurement noise covariance
R innovation covariance 0 initial condition
t time k discrete time index

the revolution. Such a mean atmospheric model from body by comparing the observed re-entry with the pre-
[6] is utilised to demonstrate the effectiveness of the dicted re-entry. These objects re-entered the Earth’s at-
proposed approach. However, more sophisticated orbit mosphere on 4th March 2000, 12th July 2001 and 19th
propagators with high fidelity atmospheric models like January 2002, respectively.
MSIS-86 [7] can be implemented with the present pro-
cedure to make the re-entry prediction more accurate 2. Filter world scenario: state equations
for high end users.
Though the Inter Agency Debris Co-ordination Com- The state equations governing the variables (a, e, B),
mittee (IADC) conducts many re-entry campaigns the are
procedures that have been adopted by the various groups ak = ak−1 + ak−1 + w1
are not available in the open literature. The present tech-
nique for the re-entry prediction uses the Kalman fil- = ak−1 + 1 (ak−1 , ek−1 )Bk−1 + w1 , (1)
ter approach with constant gains [1,8]. Gelb [9] and ek = ek−1 + ek−1 + w2
Sugiyama [10] consider constant gain Kalman filter ap-
proaches but their method of tuning the desired filter = ek−1 + 2 (ak−1 , ek−1 )Bk−1 + w2 , (2)
parameters is manual as against the present rational pro- Bk = Bk−1 + w3 , (3)
cedure based on a suitable normalised innovation cost
function as in Anilkumar et al. [11]. where w1 , w2 and w3 denote the process noise. The
In the present problem the filter states are the semi- functional forms for 1 and 2 are the expressions given
major axis, eccentricity and the ballistic coefficient (B = in King-Hele [4] which depend on a, e and B. The suffix
CD ×A/m) denoted by (a, e, B), and the measurements ‘k’ denotes the time instant. The detailed expressions
are the apogee height and perigee height (equivalent for 1 and 2 are provided in Appendix A.
to semi-major axis and eccentricity) derived from the
3. Filter world scenario: measurement equations
TLEs provided by agencies like USSPACECOM. The
constant Kalman gains as chosen in the present proce-
The measurements are available in terms of the orbital
dure are able to account for both the unmodelled and
parameters ‘a’ and ‘e’, in the tracked TLEs. But in the
unmodellable state and measurement errors. In this pa-
filter the predicted apogee and perigee heights in terms
per we validate our present approach based on three
of the above are
typical re-entries of debris objects, namely, US Sat. No.
25947, SROSS-C2 Satellite and COSMOS 1043 rocket (h−
a )k = ak (1 + ek )
A.K. Anilkumar et al. / Acta Astronautica 61 (2007) 831 – 839 833

and function J fundamental to the Kalman filter as suggested


by Sorenson [12] is defined as in [13,14]
(h−
p )k = ak (1 − ek ). (4) 
J = (1/N ) Tk [HPHT + R]−1 k
The measurements at time tk are of the form 
= (1/N ) Tk [R]−1 k, (8)
(ha )k = (h−
a )k + v 1
where the innovation k = (measurement–model
and output), and R is the covariance matrix of the innova-
tion written as HPHT +R, where H is the measurement
(hp )k = (h−
p )k + v 2 , (5)
matrix, relating the state to the measurement as
where v1 , v2 are, respectively, the measurement noises Z = HX. (9)
in apogee height and perigee height. Since no infor-
mation is available regarding the characteristics of the P is the state covariance and R is the measurement
above measurement noises in apogee height and perigee noise covariance matrix. Though there are many tech-
height values derived from the mean motion and eccen- niques for tuning the filter statistics, the heuristic ap-
tricity in TLE sets, we assume them to be white Gaus- proach of Myers and Tapley for Q and R [15] and of
sian distribution represented by Gemson and Ananthasayanam [14] for P0 is perhaps
the simplest. In general, R varies as a function of time
[v1 , v2 ]T ∼ N(0, R) (6) depending on the choice of P0 , Q and R.
Presently, it was found to be adequate to obtain P0 by
with R the measurement noise covariance matrix as-
utilising the deviations of the assumed initial conditions
sumed as constant.
of the states (a, e) with those obtained from the first
The predicted values of the apogee and perigee
TLE set under consideration and for the state B, the
heights and the ballistic coefficient useful in the state
deviation of initially assumed B0 with that of the derived
equations are updated by utilising the measured values
value of B from B ∗ is used. The values of B ∗ are B are
of the above heights as
related as
⎛ +⎞ ⎛ −⎞ ⎡ ⎤
ha ha k11 k12  − B ∗ = (ref /2)B, (10)
⎜ h+ ⎟ ⎜ h− ⎟ ⎢ ⎥ ha − h a
⎝ p ⎠ = ⎝ p ⎠+⎣ k21 k22 ⎦ . (7)
hp − h −
p k where ref is the atmospheric density at a reference
B k+ −
B k k31 k32 altitude.
The superscript (+) corresponds to the updated values, However, for Q and R we have utilised the heuristic
the superscript (−) for the predicted values and suffix estimators of Myers and Tapley [15]. These estimators
k denotes the time instant. are constructed from the statistics of the state and mea-
surement samples generated in the Kalman filter algo-
rithm. The present approximate structure for the noise
4. Adaptive Kalman filtering approach with time
statistics in one filter pass through the data are given
varying gains
below.
Measurement noise covariance
Before applying the constant Kalman gain approach
N
extensive studies have been conducted using the adap- 1 
tive technique. It turned out that generally after a tran- r̂ = rk ,
N
sient the Kalman gain matrix elements tend to constant k=1
values. This observation provides a possible approach N
1 
in which instead of tuning the usual Kalman filter statis- R̂ = [(rk − r̂)(rk − r̂)T ]. (11)
tics for the initial state, process and measurement noise N −1
k=1
covariances P0 , Q and R, a smaller number of Kalman State noise covariance
gain elements can be worked out. A good statistical con-
N
sistency of the results from the adaptive technique and 1 
q̂ = qk ,
the constant Kalman gain approach provided the confi- N
k=1
dence in the latter approach.
Assuming that the measurements are available at N 1 
Q̂ = [(qk − q̂)(qk − q̂)T ], (12)
discrete time instants, the normalised innovation cost N −1
834 A.K. Anilkumar et al. / Acta Astronautica 61 (2007) 831 – 839

where Table 1
Parameters utilised in the implementation of GA for the re-entry
rk = Zk − Hk X̂k−1 , problem

qk = X̂k − X̂k−1 (13) Parameter Values


are the state and observation samples,  is the state Population size 100
transition matrix, H is the measurement matrix, X and Bit length 20
X̂ are, respectively, the predicted and updated state Probability of cross over 0.90
estimates. Further details of iteratively obtaining all Probability of mutation 0.05
Number of generations for convergence 50
the filter statistics are available in Gemson and Anan- Alternate tolerance for convergence: change in J 0.0001
thasayanam [14]. between generations

5. Constant Kalman gain approach

Generally after the initial transients the Kalman Gain Step 6: Evaluate the strings of the new population.
matrix tends to a constant value as has been noticed Step 7: T = T + 1, go to step 5.
in the rendezvous and docking [16], evolution and ex- Table 1 provides the parameter values for the present
pansion of the space debris scenario [11] problems, it study arrived after some trials.
was also noticed in the present case as well. These con-
stant gains can be obtained by making the above nor- 7. Performance of the algorithm in different case
malised innovation cost function equal to the number studies
of measurements in the present algorithm to handle the
re-entry problem. Further, it is assumed that the above This section provides the results obtained by the
R is a constant as in [17]. Thus, the R can be estimated present algorithm for three different re-entry objects.
similar to the estimation of the pure measurement noise
only as is the case in the Method of Maximum Likeli- 7.1. Case study with US Sat. No. 25947, Soyuz
hood Estimation [17]. For this purpose we have utilised
the Genetic Algorithm [18,19] in the present work. The Sat. No. 25947, with international designation
There could be slight differences in the gain values 1999-058E, is a rocket body (Starsem Soyuz—Ikar third
obtained from the adaptive and the constant gain ap- stage, block 1 of Soyuz 11A511 U launcher), launched
proaches due to the relative periods of transients and on 18th October 1999. Its mass is about 2300 kg and is
the steady state conditions. The advantage of the con- nearly cylindrical in shape, with end domes, of 2.7 m
stant Kalman gain approach based on utilising the cost diameter and 8.1 m long. A set of 72 orbital elements
function J is that one need not propagate the covariance was made available for re-entry prediction between 2nd
equations at all, thus enormously reducing the compu- February and 3rd March 2000. The actual re-entry of
tational load. this body occurred on 4th March 2000 at 5 h 50 min,
64.243 days calculated from 1st January 2000.
6. Genetic algorithm In all the prediction studies, the object was assumed
to re-enter the Earth’s atmosphere when its semi-major
The steps to implement the genetic algorithm [18,19] axis is below 120 km above the Earth. This is because
are as follows. diffusive equilibrium predominates above 120 km with
Step 1: Choose a suitable coding to represent the significant variations in the atmospheric properties with
unknown parameters in the problem. solar, magnetic activity and local time [20].
Step 2: Select the criteria for Reproduction (Mating Starting from the 22nd TLE set, the present algorithm
pool generation)/Cross Over/Mutation. Input the initial is utilised for the prediction of the re-entry time. The
population size, probabilities of cross over and muta- present algorithm utilises six constant Kalman gains
tion, search domain of the variables, termination criteria two each for the three states, namely, apogee height,
(TC) or Maximum allowable iterations (Tmax ). perigee height and ballistic coefficient with respect to
Step 3: Set T = 0. the apogee and perigee height measurements. An im-
Step 4: Evaluate each string of the population for portant parameter in this implementation is the initial
fitness J. assumed value of the ballistic coefficient (B0 ).
Step 5: If T > Tmax or satisfies TC: terminate perform Starting with different B0 values, the performance of
reproduction, cross over, mutation on the population. the algorithm is provided in Fig. 1 by considering a set
A.K. Anilkumar et al. / Acta Astronautica 61 (2007) 831 – 839 835

Fig. 3. Re-entry prediction with different sets of Kalman gains (Sat.


Fig. 1. Re-entry predictions with different B0 values (Sat. No.
No. 25947).
25947).

effect that too based on mean atmospheric conditions.


The uncertainty in such an environmental model is also
contained in the predicted ballistic coefficient. Fig. 2
clearly brings out the robustness of the present constant
Kalman gain approach though beginning with different
values of B0 the ballistic coefficients seem to converge
towards a single value after filtering through some TLE
measurements.
The procedure for fixing the approximate B0 can be
extended to obtain the set of constant Kalman gains as
well. Fig. 3 provides the performance of the algorithm
with different sets of Kalman gains. This figure also
indicates very subtly the robustness of the filter per-
formance in terms of different sets of constant Kalman
gains. Such a variation of the Kalman filter statistics
would have led to much larger performance variations.
Fig. 2. Estimates of time varying B with different B0 values (Sat. A careful study of the normalised innovation cost func-
No. 25947).
tion for each set provided a set of constant gains as 0.6,
0.2, 0.2, 0.6, 0.00014 and 0.0001, as just about the best
of constant values for Kalman gains. From this figure it possible choice. Hence, a final prediction is made on
is clear that the value 0.040 as B0 provides the minimum the re-entry of Sat. No. 25947 with the above Kalman
variance in the re-entry prediction time with respect to gains and with B0 equal to 0.04. A graphical represen-
the TLEs. tation of this final prediction is provided in Fig. 4 and
Fig. 2 provides the ballistic coefficient update in each the other details of the results are as given below.
of the above cases as the available TLE set numbers Near optimal B0 = 0.040.
increase. As already discussed, the ballistic coefficient Actual re-entry
as a state gets updated online. This is to be expected
since starting from an initial value the filter estimates = 64.243 days from 1st January 2000
are improved, and also it accounts for the effect of tum- = 4th March 2000 at 5 h 50 min .
bling motion, and with varying gas–surface interaction.
Near optimal Kalman gains, K
Also, it may be noted that the model utilised here for
the propagation of the orbital characteristics is a very = [k11 k12 k21 k22 k31 k32]
simple one, which considers only the atmospheric drag = [0.6 0.2 0.2 0.6 0.00014 0.0001].
836 A.K. Anilkumar et al. / Acta Astronautica 61 (2007) 831 – 839

Fig. 4. Final prediction of re-entry time with near optimal Kalman


gains (Sat. No. 25947). Fig. 5. The B and cost J for prediction with optimal gains (Sat. No.
25947).

Prediction with latest elset


= 64.3257 days from 1st January 2000
= 4th March 2000 at 7 h 35 min .

The prediction depends on the epoch at which the


last TLE is available at which the re-entry is predicted.
In principle, the optimal gains can be found out as and
when the data are available. The present approach is
exercised from TLE set number 22 (about 12 days prior
to actual re-entry) to the last available TLE by predicting
the re-entry. The summary of the results is

mean prediction
= 64.2323 days from 1st January 2000
= 4th March 2000 at 5 h 35 min ;

RSS error on prediction from actual = 1 h 51 min; Fig. 6. Sat. No. 25947 re-entry predictions: measurement noise
standard deviation on prediction = 2 h 09 min. estimates in apogee and perigee heights from the adaptive filtering
technique.
Further the noise in the measurements and the states
are estimated using the adaptive procedure outlined
above based on the cost function J. The ballistic co-
efficient estimated sequentially and the corresponding 7.1.1. Remarks on statistical behaviour of genetic
cost J are provided in Fig. 5. The Fig. 6 brings out algorithm populations
the estimated measurement noise levels by adaptive The statistical behaviour of the constant gains, tuned
estimation of the noises. The estimated process noise through genetic algorithm, was also studied. It is found
values for the three states namely the semi-major axis, that even after 100 iterations the populations have not
eccentricity and the ballistic coefficient are given in converged close to a single set but there was a scatter.
Fig. 7. The Fig. 8 provides the measurement updates But the dispersion band of the parameters had narrowed
with the observations in the present approach. down as the iteration progresses.
A.K. Anilkumar et al. / Acta Astronautica 61 (2007) 831 – 839 837

reckoned from 1st January 2001. A total of 28 TLE sets


were available for the present analysis.
Starting from the second TLE set, the present algo-
rithm is utilised for prediction of the re-entry time. As
in the case of Sat. No. 25947, here also a detailed study
was carried out on the sensitivities of B0 parameter and
the Kalman gains. The details of the results are as given
below.
Near optimal B0 = 0.014.

Actual re-entry
= 193.129 days from 1st January 2001
= 12th July 2001 at 3 h 06 min .

Near optimal Kalman gains, K


= [k11 k12 k21 k22 k31 k32]
= [0.5 0.2 0.2 0.5 0.0006 0.0006].
Fig. 7. Sat. No. 25947 re-entry prediction: process noise estimates
in a, e and B from the adaptive filtering technique.
Prediction with latest elset
= 193.1488 days from 1st January 2001
= 12th July 2001 at 3 h 36 min .

7.3. Case study with COSMOS 1043 object

This is a Russian Vostok (SL-3/A-1) Upper Stage


1043 R (USSPACECOM Catalog No. 11056), started
on 10th January 2002. The object was of cylindrical
shape (diameter 2.6 m, length 3.0 m) having a dry mass
of 1400 kg. The COSPAR ID No. of this rocket body is
1978-094B. It deployed its payload (COSMOS 1043)
into an initial orbit of 622 km × 635 km at 81.2◦ incli-
nation on 10th October 1978. Fifty TLE sets for this ob-
ject were available for offline predictions, which were
downloaded from the Internet.
As in the earlier cases, the present algorithm is
utilised for prediction of the re-entry time. A complete
analysis was made on the sensitivities of B0 parameter
Fig. 8. Observed measurements and estimated values of apogee and and the Kalman gains. The details of the results are as
perigee heights from the constant gain approach. given below.
Near optimal B0 = 0.010.
7.2. Case study with SROSS-C2 Satellite Actual re-entry

Sat. No. 23099 is India’s SROSS-C2 Satellite (inter- = 19.923 days from 1st January 2002
national designation 1994-027A) launched on 4th May = 19th January 2002 at 22 h 09 min .
1994 initially into an orbit of 430 × 620 km size at
nearly 46◦ inclination. Its weight is about 113 kg and Near optimal Kalman gains, K
it is cylindrical in shape with a diameter of 0.32 m and
= [k11 k12 k21 k22 k31 k32]
length of 0.86 m. It is understood that its re-entry has
occurred on 12th July 2001 at 3 h 6 min—193.129 days = [0.6 0.2 0.2 0.6 0.0004 0.0004].
838 A.K. Anilkumar et al. / Acta Astronautica 61 (2007) 831 – 839

Prediction with latest elset The orbital parameters are now liable to change. By
= 19.942 days from 1st January 2002 differentiating the energy equations and introducing the
drag forces as the perturbation force f, we obtain fol-
= 19th January 2002 at 22 h 35 min. lowing King-Hele [4]:
 2
8. Conclusions (1 + e cos E)3/2
a = −a  2
 dE
0 (1 − e cos E)1/2
A novel approach by utilising the Kalman filter for- and
malism with constant gains for an efficient online pre-  2  
diction of the re-entry time of space objects has been 1 + e cos E 1/2
x = −a  2
(cos E + e) dE,
presented. Illustrative examples, which show the effi- 0 1 − e cos E
ciency of the proposed methodology with actual re-
where x = ae.
entries of three space objects, are also provided. This
Expanding the above in a power series of e cos(E)
methodology utilised only a mean atmosphere and a
(for e < 0.2), we obtain modified Bessel functions Jj
simple propagator considering only atmospheric drag
of the first kind of order j, defined as
effects. In particular, even the atmosphere characteris-
 2
tics as in the latest MSIS-86 model [7] can also be used. 1
However, the robustness of the constant gains have the Jj (c) = cos(j ) exp(c cos()) d.
2 0
ability to handle this mismatch as well, just as the inac-
curacies in modelling B, as well as both the unmodelled Hence utilising the modified Bessel functions and a
and unmodellable state and measurement noise charac- simple model for air density, assuming the density
teristics. depends solely on the distance r from the Earth’s cen-
tre and varies exponentially with r, we get the rate of
change over one revolution in the elements due only to
Appendix A
drag as
A.1. Basic equations of orbital decay due to air drag arev = − 2a 2 p
⎧ 3e2 ⎫
Consider an object S moving due to an attractive ⎪
⎨ J0 + 2eJ 1 + (J0 + J2 ) ⎪

4
force which varies inversely with the square of the dis- × exp−c
tance r from the centre of attraction. If the position vec- ⎪
⎩ e 3 ⎪

+ (3J1 + J3 ) + O(e4 )
tor of the particle relative to the centre of attraction is 4
r, we have erev
r = −2ap
r̈ + =0
r3 ⎧ ⎫
⎪ e e2
and the angular momentum h is given by ⎨ J1 + (J0 + J2 ) − (5J1 + J3 ) ⎪ ⎬
2 8
× exp−c ,
h = (p)1/2 n, ⎪
⎩ e3 ⎪

− (5J0 + 4J2 − J4 ) + O(e4 )
16
where ‘n’ is the unit vector perpendicular to the orbit
where p is the density at perigee, c = ae/H , H
plane and ‘p’ is the semi-latus rectum defined as
is the density scale height,  = (QC D )A/m, Q the
p = a(1 − e2 ) = h/2 . factor which includes the rotation of the atmosphere
(0.9 < Q < 1.1), m the mass of the satellite, Jj the
The energy equation is modified Bessel function of first kind of order j , in all
  cases having argument c, CD the drag coefficient usu-
1 1 1
ṙ • ṙ =  − , ally taken as ≈ 2.2, and A the effective reference area.
2 |r| 2a Hence at kth revolution
where the dot (·) denotes the scalar product. ak = ak−1 + ak−1 ,
Suppose now that the orbit is perturbed by a force f
per unit mass, so that ek = ek−1 + ek−1 ,

r where aj , ej stands for semi-major axis and eccentricity


r̈ + = f. at time instant tj , j = k − 1, k.
r3
A.K. Anilkumar et al. / Acta Astronautica 61 (2007) 831 – 839 839

A.2. The model for the density variation in the [8] A.K. Anilkumar, M.R. Ananthasayanam, P.V. Subba Rao,
atmosphere Prediction of Re-entry of Space Debris Objects: Constant Gain
Kalman Filter Approach, AIAA-2003-5393, 2003.
[9] A. Gelb (Ed.), Applied Optimal Estimation, MIT Press,
The model of the atmosphere used here assumes Cambridge, MA, 1974.
a spherically symmetrical distribution of particles in [10] N. Sugiyama, System identification of jet engines, Journal of
which the density  varies exponentially according to Engineering for Gas Turbines and Power 122 (2000) 19–26.
  [11] A.K. Anilkumar, M.R. Ananthasayanam, P.V. Subba Rao,
(h − h0 ) New approach for the evolution and expansion of space
 = 0 exp − , debris scenario, Journal of Spacecraft and Rockets 43 (2006)
H
1271–1282.
[12] H.W. Sorenson, Least squares estimation from Gauss to
where h is the height above the surface of the Earth, 0
Kalman, IEEE Spectrum 7 (1970) 63–68.
the density at height h0 , and H the scale height. [13] R.M.O. Gemson, Estimation of aircraft aerodynamic derivatives
accounting for measurement and process noise by EKF through
adaptive filter tuning, Ph.D. Thesis, Department of Aerospace
References Engineering, Indian Institute of Science, Bangalore, India, 1991.
[14] R.M.O. Gemson, M.R. Ananthasayanam, Importance of Initial
[1] A.K. Anilkumar, New perspectives for analyzing the breakup, State Covariance Matrix for the Parameter Estimation using
environment, evolution, collision risk and reentry of space Adaptive Extended Kalman Filter, AIAA-98-4153, 1998.
debris objects, Ph.D. Thesis, http://etd.ncsi.iisc.ernet.in/handle/ [15] K.A. Myers, B.D. Tapley, Adaptive sequential estimation with
2005/80, Department of Aerospace Engineering, Indian unknown noise statistics, IEEE Transactions On Automatic
Institute of Science, Bangalore, India, 2004. Control 21 (1976) 520–525.
[2] N.L. Johnson, D.S. McKnight, Artificial Space Debris, Orbit [16] N.K. Philip, M.R. Ananthasayanam, Relative position and
Book Company, 1987. attitude estimation and control schemes for the final phase of an
[3] Web sites for two line element sets http://www.geocities.com/ autonomous docking mission of spacecraft, Acta Astronautica
iss25544/#TLEs or www.wingar.demon.co.uk or http:// 52 (2003) 511–522.
www.celestrak.com. [17] N.K. Philip, M.R. Ananthasayanam, The method of maximum
[4] D. King-Hele, Satellite Orbits in an Atmosphere: Theory and likelihood estimation: a generalized computer program, Report
Applications, Blackie, London, 1987. No. 95 FM 3, Department of Aerospace Engineering, Indian
[5] D.A. Vallado, Fundamentals of Astrodynamics and Appli- Institute of Science, Bangalore, 1995.
cations, Second ed., Microcosm Press, Kluwer Academic [18] L.D. Whitley (Ed.), Foundations of Genetic Algorithms, Morgan
Publishers, Dordrecht, 2001. Kaufmann, Los Altos, CA, 1993.
[6] US Standard Atmosphere, NOAA, NASA and USAF, U.S. [19] K. Deb, Optimization for Engineering Design: Algorithms and
Government Printing Offices, Washington, DC, 1976. Examples, Prentice-Hall of India, 1995.
[7] A.E. Hedin, MSIS-86 thermospheric model, Journal of [20] R.C. Whitten, W.W. Vaughan, Guide to Reference and Standard
Geophysical Research 92 (1987) 4649–4662. Atmosphere Models, AIAA-G-003-1988, 1990.

You might also like