INVESTIGATION OF SIMPLIFIED MODELS FOR ORBIT DETERMINATION USING

SINGLE FREQUENCY GPS MEASUREMENTS

Ana Paula Marins Chiaradia*

Hélio Koiti Kuga
Antonio Fernando Bertachini de Almeida Prado
INPE – Instituto Nacional de Pesquisas Espaciais
Av. dos Astronautas, 1758 - Jardim da Granja
São José dos Campos - SP - Brazil - CEP: 12 227-010
*[email protected]

Abstract position observable geometrically. If the measurements

are accurate, a sequentially dynamical orbit
The main goal of this work is to investigate simplified determination may not need such a precise force model
models to determine in real time the orbit of an artificial as the whole information is locally provided by the
satellite, using single frequency GPS measurements. measurements. This approach was applied in the work
This model should be compact providing standard of Wu et al.11, which uses the so-called reduced
precision at low cost. Cowell's method has been used to dynamic technique.
propagate the orbit state vector. The modeled forces are The state estimation method adopted was the Kalman
geopotential up to 23rd order and degree of the spherical filter which is used to estimate the spacecraft's orbit on
harmonic coefficients. To propagate the state covariance board, excluding the need of iterating the data collected
matrix, it has been considered a more simplified model previously and being able to provide the current orbit in
than the one used in dynamical model. For computing real time.
the state transition matrix, it is considered only
keplerian motion. To estimate an orbit in real time the
extended Kalman filter has been used throughout this Estimation Method
paper. Many tests are carried out starting from the
simplest two body model and varying the contribution The orbit determination problem, which has a non-
sources of the errors to be considered. In order to assess linear dynamical system and a non-linear measurement
the results, the estimated orbit for several cases is system, can be formulated in a way that it is possible to
compared with a full reference Kalman filter, for the apply one of the best known methods of sequential
Topex/Poseidon satellite. linear estimation, the Kalman Filter.
The extended Kalman filter is a Kalman filter version
Key words: Kalman Filtering, GPS, Real time orbit applicable to problems like this one, constituted by a
determination. time-update and a measurement-update cycles. The
time-update phase updates the state and the covariance
matrix along the time using the dynamical equations.
Introduction In this work, a simple reduced state vector is chosen to
be estimated:
The GPS system has been used to determine artificial
satellite orbits because it may provide orbit
determination precision as good as or better than
(
x = r T ,vT , b , ) (1)
methods using conventional tracking stations. The later
provides standard precision around hundred meters and where r T = (x, y, z) and v T = ( x, y, z ) are the
the former can provide precision as tight as some spacecraft's position and velocity vector, and b is the
centimeters. The attractiveness of GPS is magnified receiver clock offset.
because of lower costs and autonomous navigation Then, the differential dynamic equations of motion to
resources. With the advances of technology, the single be integrated are given by:
frequency GPS receivers provide a good basis to
achieve enough precision at relatively low cost still x = f ( x, t ) , (2)
attaining the accuracy requirements of the mission.
The GPS system provides at a given instant a set of
many redundant measurements which makes the orbit and
 Φ = FΦ , (3)
where Yk is the observed measurement.
where f ( x , t ) is the vector-valued function of time and The measurement update phase uses the Kalman
equations to incorporate the information given by the
the state, Φ is the state transition matrix which relates
measurements themselves, and obtains improved
the state between tk and tk+1, and F = ∂f ( x, t ) / ∂x . estimates of the state and of the covariance:
Both equations should be numerically integrated
simultaneously, so that F is evaluated always along the
most current state x . Next, one updates the covariance (
K k = Pk H kT H k Pk H kT + Rk )
−1
, (9)
matrix P by means of the discrete Riccatti equation:
xˆ k = xk + K k yk , (10)
Pk +1 = Φ k +1PˆkΦ kT+1 + Qk , (4)
Pˆk = (I − K k H k )Pk , (11)
where P̂k is the covariance matrix after processing all
measurements at time tk, Φk is the state transition matrix where Rk is the discrete measurement noise covariance
obtained by the previous integration and Qk is the which is basically a measurement weight matrix. These
discrete state-noise covariance which is a measure of the equations can be used to process measurements
error between the reference state and true state arising sequentially so that the matrix inversion in (9) is a
from imperfect modeling. The P matrix is a measure of scalar. To be precise, the measurements should be
how well the errors are known. uncorrelated, in which case the measurement covariance
noise Rk is diagonal.
At the end of this process xk +1 and Pk +1 , are
obtained and are called time-updated state and
covariance, respectively. Dynamical Model
The measurement residual and sensitivity matrix are
found by forming the computed observation equation. The choice of the propagation method depends on the
The model for a GPS pseudorange measurement is requirements of the mission. As the goal of this work is
given by: to have relatively standard accuracy along with
minimum computational cost, the Cowell's method has
yc ( x k , t k ) = ρ k + β k , (5) been used to propagate the state vector.
The modeled forces in this work are due to
where geopotential taking into account the spherical harmonic
coefficients up to 23rd order and degree of GEM10B
model. The integration is carried out by using the simple
ρ = ( xGPS − x) 2 + ( yGPS − y ) 2 + ( zGPS − z ) 2 (6) fourth order Runge-Kutta algorithm without any
mechanism of step adjustment or error control. The
is the geometric range, x, y, and z are the positional fourth order Runge Kutta is considered an adequate
states of the user satellite at reception time, xGPS, yGPS, numerical integrator due to its simplicity, fair accuracy,
and zGPS are the positional states of the GPS satellite at and low computational burden.
transmission time (corrected for light time delay), and The dynamic equation of motion is given by:
βk are errors steaming from ionosphere path delay, GPS
satellite and receiver clock offsets, GPS ephemeris µ
error, multipath, and other unmodeled errors. r =− + a GEO , (12)
r3
Using the above equation the sensitivity matrix is
given by:
where µ is the geo-gravitational constant and aGEO is the
acceleration due to perturbing geopotential, computed
 (x − x ) ( yGPS − y ) (zGPS − z ) 
H k = − GPS ,− ,− , 0, 0, 0,1 . (7) according to Pines8.
 ρ ρ ρ  The user satellite coordinates are in true equator and
equinox of date frame (ToD) and the computation of the
The measurement residual, or innovation sequence is: acceleration of the satellite due to the spherical
geopotential requires the coordinates of the satellite in
yk = Yk − yc ( xk , tk ) , (8) the rotating Earth-fixed equator and prime meridian
frame (EF). Then, it is necessary to provide the software
with the coordinate transformations considering the the orbit, and E is the eccentric anomaly. The
effects of precession, nutation, polar motion, and polynomial coefficients af0, af1, and af2 are transmitted
sidereal rotation. Polar motion and difference between in units of sec, sec/sec, and sec/sec2, respectively. The
UTC and UT1 was disregarded as it requires uploads of clock data reference time toc is also broadcast. The value
these parameters to the satellite computer, increasing the of tsv must account for the beginning or end-of-week
ground task load and decreasing the satellite autonomy. crossovers.
To decrease the computational burden further, The user clock offset is part of the estimated state
precession and nutation parameters and matrices may be vector.
considered constant for a span of one day.

Ionospheric Correction
Measurement Model
The GPS transmitted signals pass through the
The equation of the pseudorange in L1 frequency is ionospheric layer causing errors in the measurements,
given by: being the ionospheric effects one of the hardest effects
to model.
Pk1 = ρ k + I k + c[dt GPS (t k ) − dtU (t k )] − ε k , (13) As the user is a satellite, tropospheric effects are
considered irrelevant, and a simple elevation mask will
reject measurements reflecting into the Earth’s
where Pk1 is the pseudorange in L1, ρk is the geometric troposphere.
range given by Eq. (6), Ik is the ionospheric delay, c is The ionospheric effects depend on the frequency. It is
the vacuum speed of light, dtGPS(tk) is the GPS satellite possible to remove them easily by using dual frequency.
clock offset, dtU(tk) is the receiver clock offset, tk is the However, in single frequency measurements, they cause
observation instant in GPS time, and εk is a remnant many real troubles in the data processing, where the
error supposed random gaussian. degradation of precision can achieve reasonable values
depending on the solar activity and the spatial
(geomagnetic) environment. The maximum effect in
Clock Error Terms single frequency receivers is around 20m and for the
case of dual frequency it is around 4.5cm9.
The third term of the right-side is the clock bias which Despite assuming single frequency in this work, the
represents the combined clock offsets of the satellite and dual frequency model has been used to correct the
of the receiver with respect to GPS time. Each GPS ionospheric effects in the pseudo-range measurements.
satellite contributes with one unknown clock bias. The As such it allows one to analyze the impact of either
information for the GPS satellite clocks is known and neglecting or not the ionosphere effect on the orbit
transmitted via the broadcast navigation message in the estimates. In this case, this is possible because our
form of three polynomial coefficients with a reference satellite test case, TOPEX/Poseidon, has a dual
time toc. The clock correction of the GPS satellite for the frequency receiver on board. The equation to correct the
epoch tGPS is6: ionospheric effects of the pseudo-range measurement in
L1 is given by6:
∆t SV = a f0 + a f1 (t sv − toc ) + a f2 (t sv − toc )2 + ∆t R , (14)
I kP =
f 22
f 22 − f12
(P1
k )
− Pk2 , (17)
with
where I kP is the ionospheric correction to range
tsv = tGPS − ∆tsv (15)
measurements in L1 frequency, Pk1 and Pk2 are the
and pseudorange measurements in L1 and L2, respectively,
f1 is the frequency in L1 (1.575 GHz), and f2 is the
2 2 frequency in L2 (1.227 GHz).
∆t R = − aµ esinE = − x• x , (16)
2
c c2

where ∆tR is a small relativistic clock correction caused

by the orbital eccentricity e, a is the semimajor axis of
 Transition Matrix The T/P receiver can track up to 6 GPS satellites on
both frequencies if Anti-Spoofing is inactive2. This raw
The calculation of the state transition matrix presents data was recorded 10 second interval of GPS time.
one of the highest computational costs because it needs The observation data in Rinex format were obtained
the evaluation of the Jacobian matrix (partial via Internet10. The navigation data in Rinex format were
derivatives) and integration of the current variational sent by Binning2, although they also could be obtained
equations. This matrix can pose cumbersome analytical via Internet7.
expressions when using a complex force model3. The precise GPS satellite coordinates are obtained
A simple force model to state transition matrix through Internet and interpolated4 to appropriate times,
considered the keplerian motion. The state transition's providing the coordinates in EF frame.
differential equation is defined as: The initial conditions to t0 = 1sec in UTC time as of
November, 18th, 1993 are given in the Table 1.
Φ (t k , t 0 ) = F (t k )Φ (t k , t 0 ) , (18)
Table 1: Initial conditions to user satellite
where Φ (t 0 , t 0 ) ≡ I is the initial condition,
coordinates.

Initial Conditions Nov 18, 1993

 ∂r ∂r  ToD x (m) 7617202.243009592
 ∂r 0 3 x1 
∂v 0 3 x 3 ToD y (m) 1235354.688733236
 0 3 x3 
∂v ∂v
Φ (t k , t 0 ) =  0 3 x1  (19) ToD z (m) -135607.5368155133
∂r ∂v 0 3 x 3 ToD x (m/sec) -353.5738692980746
 0 3 x3 
 01 x 3 01x 3 11 x1  ToD y (m/sec) 2898.599146009871
  ToD z (m/sec) 6568.36541232146

and The initial conditions are obtained through the

TOPEX/Poseidon Precise Orbit Ephemeris (POE)10
 0 3 x3 I 3 x3 03 x1  generated by the Jet Propulsion Laboratory (JPL) in
F (t k ) = G (t k )3 x3 03 x1  .
UTC time. The JPL POE is claimed to estimate Topex's
0 3 x3 (20)
position to an accuracy of better than 15 cm (See Ref. 1
 01x3 01x3 01x1  for details). The states in the POE are provided in one
minute increments in Inertial True of Date coordinates
G (t ) = ∂f (r , t ) / ∂r is the gradient matrix of the in UTC time. But, the TOPEX GPS measurements are
in 10 second increments in GPS time. Accordingly to
geogravitational force and f(r,t) is the acceleration of
IERS, the difference between the UTC and GPS times is
the satellite that, in this case, considered the Keplerian
9 seconds to this date. Therefore, it was necessary to
motion.
interpolate the states in one second steps through our
ODEM Orbit Determination software4,5.
Data Set
Table 2: Extended Kalman filter input parameters.
As observation data TOPEX/Poseidon (T/P) GPS data
set is used, because this satellite has dual frequency Filter Input Parameter Initial values
GPS receiver onboard and such data are easily found in Pr (m2) 1.0 × 104
Internet. Pv (m2/sec2) 1.0 × 10-2
As suggested by Binning2, one uses the T/P data set of Pb (sec2) 1.0
November 18th, 1993, because Selective Availability Qr (m2) 1.0 × 101
(SA) was also not in operation for 17 of the 25 available Qv (m2/sec2) 1.0 × 10-5
GPS satellites allowing civilian users access to the most Qb (sec2) 1.0 × 10-3
precise GPS measurements. At that time the GPS
constellation was not yet considered fully operational,
Rk (m2) 2.5 × 101
and therefore, Anti-Spoofing was also off. This allowed
all users to receive clean data in both L1 and L2 The Kalman Filter input parameters are given in the
frequencies. Table 2, where Pr is the position error, Pv is the velocity
error, Pb is the user clock error, Q is the corresponding
dynamical noise, and Rk is the covariance of
measurement error. Table 4: Accuracy of the fixed step integrator RK4.
The considered constants are given in Table 3.
Case After ∆t (sec) ∆r (m) ∆v (m/sec)
Table 3: Constants in Topex case study.
1 60 sec 1 13.248 0.442
WGS 84 gravitational 3.986005 × 1014 2 hours 43594.4 40.801
parameter m3/sec2 2 60 sec 10 13.248 0.442
WGS 84 rotation rate 7.292115167 × 10-5 2 hours 43594.4 40.801
rad/sec 3 60 sec 30 13.248 0.442
WGS 84 Earth's radius 6378137 m 2 hours 43593.7 40.800
speed of light 299792458 m 4 60 sec 60 13.246 0.442
J2 1.08263 × 10-3 2 hours 43581.3 40.787
5 60 sec 1 0.121 0.004
2 hours 641.805 0.585
Results 6 60 sec 10 0.121 0.004
2 hours 641.814 0.585
All tests are accomplished considering the initial 7 60 sec 30 0.121 0.004
conditions cited in Tables 1 and 2. 2 hours 642.579 0.586
The first test is aimed at checking the accuracy of the 8 60 sec 60 0.121 0.004
integrator of fourth order Runge-Kutta. The filter is 2 hours 656.153 0.598
turned off and several fixed step sizes are attempted. 9 60 sec 1 0.012 0.
Also the geopotential truncation order is changed to 2 hours 87.601 0.084
verify the different level of errors. 10 60 sec 10 0.012 0.
In the cases 1 to 4, it is considered the Keplerian 2 hours 87.609 0.084
motion. In the cases 5 to 8, the geopotential includes the 11 60 sec 30 0.013 0.
term in J2. In the cases 9 to 12, up to 23rd order and 2 hours 88.387 0.085
degree of spherical harmonic coefficients are 12 60 sec 60 0.042 0.
considered. The step size (∆t) varies from 1 second to 1 2 hours 102.123 0.098
minute. In Table 4, ∆r and ∆v are the errors on the
position and velocity states, respectively, compared with The full model Kalman filter results are then used to
the POE. compare the several contribution sources to the errors.
Table 4 shows the maximum accumulated errors for Fig. 1 shows the pseudorange residuals in L1 during the
each case after both one minute and two hours of first 15 minutes of the filtering. Also, one can note that
integration with different step sizes. One can conclude the filtering is converging. It is impossible to show all
that better accuracies are obtained in the cases 9 and 10. residuals because of scaling. Fig. 2 shows the number of
As the difference between them is very small, the 10 GPS satellites at each epoch of the filtering. The
second step size is adopted in all foregoing cases with a average number of available GPS satellites is 5.536 and
goal of minimizing the computational cost. the average number of used GPS satellites (not rejected)
Therefore, in the remaining cases, the TOPEX's orbit is 5.515.
state is generated every 10 second step size maximum. For cases 13 and 14, the tests are related to the
The next results are achieved with the filter turned on, dynamical model, i.e., only the perturbation of the
considering around 2 hours of data. The reference dynamical model changed in each case. In case 13, the
trajectory is accomplished considering the full model pure Keplerian motion was considered, while in the case
Kalman filter, which shows good agreement with the 14, the effect of J2 is considered. Table 5 shows the
JPL/POE ephemeris. It considered up to 23rd order and results obtained compared with the reference case.
degree of the spherical harmonic coefficients and the
effects of precession and nutation. The relativistic,
ionospheric, GPS and user receiver clock offsets are
considered in the measurement model as shown
previously.
 Conclusions

30 This paper considers several aspects of modeling when

Residuals (m)

15
using the GPS for real time orbit determination. It is
observed that for sampling rates of 10 seconds and
0
0 3 6 9 12 15
using a simple fourth order Runge-Kutta numerical
-15 integrator the higher order geopotential model can yield
already reasonable results. At least J2 shall be included
-30
in the geopotential model because a single Keplerian
Time (min) model can not account for the short period oscillations,
even for 10-second step size.
Figure 1: Sample of Residuals for the full model. The measurement model shall be carefully modeled as
some effects can be very pronounced. The GPS clock
offset shall be taken into account, secondly the signal
Number of GPS

6 travel time correction. After comes the user clock offset,

the ionosphere, and to a lesser extent the relativistic
satellites

4
effect. As a last remark, the precession and nutation of
2 coordinates shall be taken into account as they
0 considerably change the actual coordinates. This leads
0 20 40 60 80 100 120 to an additional computational burden which can be
minimized if one considers the precession and nutation
Time (min) parameters constant over a given period, say one day.

Figure 2: Number of processed GPS satellites in two Table 6: Errors due to variation of the measurement
hours run. model.

Case RMSr (m) RMSv (m/sec)

Table 5: Errors due to variation of the dynamical 15 2.03×102 0.15
model. 16 3.71×101 3.03
17 6.27 0.008
Case RMSr (m) RMSv (m/sec) 18 3.86 0.005
13 2.87×101 4.21 19 6.89×101 0.09
14 0.33 0.05
Table 7: Errors due to neglecting of the effects of
In cases 15 to 18, one of the following effects are not precession and nutation.
considered in each case: GPS clock offset, receiver
(satellite) clock offset, relativistic effect, ionosphere Case RMSr (m) RMSv (m/sec)
effect, and signal travel time. To depict the relative
20 8.8×103 7.84
importance of each one to the accuracy of the orbit
estimates, in case 15 the GPS clock offset is not
considered; in case 16 without considering the receiver
clock offset; in case 17, without accounting for the Acknowledgements
relativistic effect; in case 18, without the ionospheric
correction; and, in case 19, without the signal travel The authors wish to express their appreciation for the
time correction. Table 6 shows the obtained results support provided by CAPES (Fundação para
compared with the reference case. Coordenação de Aperfeiçoamento de Pessoal de Nível
Finally, in case 20, the effects of precession and Superior) - Brazil and INPE (Brazilian Institute for
nutation in the coordinates are disregarded (See Table Space Research) - Brazil. We would like to thank
7), when transforming the coordinates from J2000.0 to Patrick W. Binning of the Naval Research Laboratory
True of Date (See Table 7). and Colorado Center for Astrodynamics Research at the
University of Colorado for providing kindly the GPS
navigation files.
 References
1
Binning, P. W. GPS, Dual Frequency, SA Free Satellite
Navigation. Proceedings of ION 52nd Annual Meeting,
p803-812, Boston, Mass., June, 1996.
2
Binning, P.W. Satellite Orbit Determination Using
GPS Pseudoranges Under SA. Advances in The
Astronautical Sciences, AAS97-111, Vol. 95, part I,
p183-193, 1997.
3
Kuga, H.K. Estimação adaptativa de órbitas aplicada a
satélites a baixa altitude. Master Thesis, INPE, São José
dos Campos, INPE-2316-TDL/079, 1982. (in
portuguese).
4
Kuga, H. K.; Gill, E. A Mathematical Description of
the ODEM Orbit Determination Software. DLR-GSOC-
IB-94-06, DLR, 1994.
5
Kuga, H. K.; Gill, E.; Montenbruck, O. Orbit
Determination and Apogee Boost Maneuver Estimation
Using UD-filtering. DLR-GSOC IB 91-2, DLR, 1991.
6
Leick, A. GPS Satellite Surveying, Second Edition,
John Wiley & Sons, INC., 1995.
7
Noll, Carey. Crustal Dynamics Data Information
System. [On-line] ftp://cddisa.gsfc.nasa.gov/, Dec.,
1998.
8
Pines, S. Uniform representation of the gravitational
potential and its derivatives. AIAA Journal, Vol.11,
No.11, Nov., 1973.
9
Seeber, G. Satellite Geodesy: Foundations, Methods,
and Applications. Walter de Gruyter, Berlim-New York,
1993.
10
Shapiro, B. R. Topex/ Poseidon Navigation Team.
[On-line] http://topexnav.jpl.nasa.gov., April, 1998.
11
Wu, S.C.; Yunck, T. P.; Thornton, C. L. Reduced
dynamic technique for precise orbit determination of
low Earth satellites. J. Guidance, Control and
Dynamics, Vol. 14, pp. 24-30, 1991.