Sei sulla pagina 1di 4

Study on the stochastic model for VRS network-based GNSS positioning

T. Jongrujinan, C. Satirapod
Department of Survey Engineering, Chulalongkorn University, Bangkok, Thailand
(Thanate.j@student.chula.ac.th, Chalermchon.s@chula.ac.th )

Abstract despite the fact that the reference stations are many
tens of kilometers apart. The main challenge is to keep
Network-based positioning has currently become on fixing ambiguities during situations when there is
the main precise GNSS surveying technique especially a long data gap, or when a new satellite rises [2].
for a medium-range baseline (approximately 20-100 Secondly, the distance-dependent errors at reference
km.). The key concept of this approach is to use the stations are precisely modelled and interpolated to a
observations of multiple reference stations to rover location within network coverage. Several
generate the network correction for mitigating the interpolation methods have been proposed and
distance-dependent error including atmospheric compared in detail [3]
effects and orbital uncertainty at the user’s location. Although the significant distance-dependent
Numerous GNSS data processing strategies have been errors can be modelled and mitigated in
adopted in the functional model in order to improve aforementioned functional models, some unmodelled
both the success of ambiguity resolution and the biases may be present. Beside the adequate functional
accuracy positioning results. However it is impossible model, In order to obtain optimal estimates from the
to completely model the aforementioned errors. As a least-square solution, the stochastic model play an
result, the unmodelled residuals still remain in the important role. The stochastic model represents the
observations when the least squares estimation is statistical characteristics of the observations in the
employed. An alternative approach to deal with these form of the variance-covariance matrix. On the other
residuals is to construct the realistic stochastic model hand, the functional model described the relationship
that GNSS observations are assumed to be and unknown estimates. The VCV matrix should be
homoscedastic. correctly defined. The standard approach to construct
This research aims to investigate stochastic models the VCV matrix is based on the variance-covariance
used for VRS technique in the rapid static mode. The propagation law. For simplicity, it is usually assumed
rigorous statistical method, MINQUE has been that all GNSS observations have the same variance,
applied to estimate the variance-covariance matrix of and they are statistically independent. In fact, these
the double-difference observations for VRS–rover assumptions are not realistic, the GNSS observables
baseline determination. The findings of the may change with different observation conditions
investigations indicated that the MINQUE stochastic such as satellite elevation. Moreover in case of VRS
model could enhance 14 percent better in the vertical technique, to do the baseline processing of the rover
accuracy but produce comparable result in horizontal observables relative to the VRS observables, the
accuracy compared to the model that use equal weight precision of the VRS observables should follow from
for all observation. propagating the uncertainty of both network error
modelling and the interpolate method which form the
Keywords: VRS, Stochastic model. VRS observables.
In this paper, the MINQUE procedure is utilized
1. Introduction for VRS technique in the rapid static mode. This
procedure determines the VCV matrix by using the
For centimeters accuracy level, network–based least squares residuals as the quality indicators. Test
positioning provides a solution the can overcome the results will be presented to evaluate the performance
constraint of the single-based positioning that the of the method.
distance between the reference station and the rover
receiver is typically not exceed 10 km due to the 2. VRS technique
presence of atmospheric biases and orbit errors. [1]
confirmed that the ionospheric error is the main error The core of network-based positioning system is to
source which degrades the network-based positioning use multiple reference stations in order to interpolate
performance in Thailand The functional model to the distance-dependent errors for specific rovers
generate network corrections is carried out in two operating in the network coverage area as well as
major steps. Integer ambiguities between reference transmit the correction information to rovers.
stations must be first correctly resolved in real-time Currently, there exists no standardized data format for
broadcasting of network correction parameter and due to insufficient knowledge about the physical
also no agreement in parameterization of correction phenomena of error sources. Therefore, high accuracy
model. There are several implementations of results using least squares adjustment also requires the
distributing correction that system could provide, reasonable selection of the stochastic model
however the virtual reference station (VRS) is the MINQUE (Minimum Norn Quadratic Unbiased
most commonly used method. It is based on the Estimation) is one of the rigorous statistical method
transmission of VRS observables of “virtual reference for estimating VCV components that all elements in
station”. The VRS is a non-existent GNSS reference the VCV matrix will be directly estimated as the
station at the approximation rover’s location, however, unknown parameter in the least square solution
VRS observables is computed from real GNSS similar to the parameters in the functional model. The
observables of reference stations resemble to the basis of this variance component estimation is to use
distance-dependent error corrections. Meanwhile, the the least squares residuals to formulate the realistic
whole computations to generate VRS are operated by VCV matrix (all GNSS observables are statistically
server side, the rovers just treat or process the received dependent and have different variances). It is assumed
VRS as the same as that of single base station. This that these residual series could adequately reflect the
may save the cost for upgrade any software and presence of the residual biases and measurement noise
receiver. In order to generate the VRS, the following due to the imperfect functional model if redundant
server- side process are required observations are available.
1. Network ambiguity: the purpose of this step is to Previous research [4] has efficiently used this
resolve the integer ambiguities for the individual procedure for the single-based positioning in static
baselines among the reference stations. mode. Therefore it’s expected that MINQUE could
2. Error Interpolation: The residual vectors from construct the appropriate stochastic model for the
step 1 are used to be inputs for interpolating the network-based positioning in rapid static mode (about
distance-dependent error at the approximate rover 15-20 epochs per session) if the observation period is
coordinate long enough to remove the most of systematic effects.
3. VRS observables generation: VRS observables The MINQUE algorithm is expressed as below:
is generated based on the interpolated errors and In the following Gauss-Markov model with n
reference station observables measurements and t unknown;

3. Stochastic model l = Ax + v

After the generated VRS observables is For a session solution with t epochs of data, the
transmitted to the rover receiver via a communication full VCV matrix of double differenced observations is

= P-1 = ∑ θT
link, then the rest of the processes on the rover side
are the same as single-based positioning. The double C (1)
differenced (DD) carrier phase observations are
widely used to formulate the functional model of least
square since it efficiently eliminated most of where l and v are n x 1 vectors of the measurements
systematic errors. The DD equation can be written as: A is the n x t design matrix; x is the t x 1 vector of the
unknown parameters; θ^1 ,θ^2 ,….,θ^k are the variance-
∅ =ρ + N −I + T + O
, , , , , , covariance components; T1, T2 ,…,Tk are the so-called
, , , , , ,
accompanying matrices ; P is the weight matrix of the
where observations and C is the VCV matrix of the
observations. A minimum norm quadratic unbiased
θ ;
u and v denote the rover station and virtual
reference station respectively; i and j denote the estimation of the linear function of
tracked satellites ; ∅ , denote the DD observable;
, g1θ1+g2θ2+…+ gkθk is the quadratic function.

ρ ,, denote the DD geometric distance; N , denote


,
Tr {M.C.M.C}=min; MA=0
the DD ambiguity ; I ,
,
,
denote the DD of
ionospheric errors, T , denote the DD of tropospheric
, where Tr{} is the trace operator of a matrix; the

errors, , O , denote the DD of orbital error.


,
variance-covariance components can be estimated as

θ^ = (θ^1 ,θ^2 ,….,θ^k ) = S-1q (2)


It is noted that that the VRS station and the rover
station are very close to each other. The DD of the where the matrix S = {Sij} with Sij = Tr{RTiRTj} (3)
distance –dependent errors will be very small and can and the vector q = {qi}
be ignored. with qi = lTRTiRl (4)
The functional model, DD, still remain the
unmodelled systematic error and observation noise and
R = PQvP (5)

with Qv = P-1 – A(AtPA)-1AT being the adjusted


residual cofactor matrix.

R can be expressed by a partitioned matrix:

R11 ⋯ R1m
R= ⋮ ⋱ ⋮
Rm1 ⋯ Rmm

where m is the number of the observation epoch in a


session
Since the relationships between v and l are:

v = -QvPl (6)
and
PQvPv = -PQvPl = Pv (7)

According to Equation (5) and (6), Equation (3) can


be further written Figure 1: Location of CORS used for experiment
qi = lTRTiRl = vTRTiRv = vTPTiPv (8) Establishment of reference coordinates
In order to obtain accurate reference coordinates of
It can be noticed from equation (2), (3), (8) that all stations, 7-day GPS data (observed during
elements of the VCV matrix are implicitly defined. February 1-7, 2010) were processed using the Bernese
Therefore an iterative process must be performed to GPS Processing Software. A one-day solution was
solve the equation. Initially, the priori value of
variance-covariance components (θ0i ) could be given
obtained for each station, using a standard processing
baseline procedure. An International GNSS service
by the standard method and the first-iteration (IGS) station was selected as a fixed station and the
estimate (θ1i ) can be obtained by equation (2). IGS final orbits were also introduced in the data
Subsequently, using the previous estimate (θ^j ) as the processing step. The weekly averaged coordinates of
priori value, the new estimate (θ^j+1 ) from the (j+1)
an individual station were used as reference
coordinates for subsequent analysis.
iteration is :
Processing strategies
θ^j+1 = S-1(θ^j )q(θ^j ) ; j=1,2,… As shown in Figure 1, one reference station
(CUSV) was used as rover station and the other five
This is called the iterated MINQUE. If θ^ converges, stations were simulated as network reference stations
the converging value of θ^ satisfy the following to generate VRS in form of RINEX by the Trimble
condition : Total Control software. Then both virtual rinex and
S(θ^ ) θ^ = q(θ^ ) rover observables rinex was cut into segments. The
various segment size is from 14 epochs to 20 epochs.
which can be further expressed as Subsequently, each data segment was separately
processed. The baseline solution between VRS
Tr{R(θ^ )Ti} = lTR(θ^ ))TiR(θ^ )l observables and rover observables was obtained by a
Matlab-based CUGPS Baseline processing software.
4. Experiment First, Float solution is carried out by least square
adjustment with DD equation as the observation
Test data equation. Meanwhile, the VCV matrix is determined
The 3-hour GPS dataset collected from six CORS by the two stochastic model (1.MINQUE procedure,
stations of Thai organizations on February 15 2010, 2. Standard procedure). Secondly, the LAMBDA
during the period from 3:00 to 6:00 AM was used to method [5] will be used to determine the integer value
test the proposed procedure. It was sampled at a 30- for the ambiguities. LAMBDA takes the estimated
second rate and the cut-off elevation angle was set to float solution together with VCV matrix as input, and
10 degree. deliver the estimated integer ambiguities. Finally, the
fix solution is computed.
5. Analysis of results observables
After all data segments (140 sessions) were
processed, the coordinate results were compared with
the reference coordinates. The performance of the
VRS technique under the MINQUE stochastic model
is evaluated by Root Mean Square Error (RMSE). The
Equal-weight stochastic model (Standard method) is
used as the comparison benchmark as shown in
Table1. As the results, MINQUE tend to produce
slightly better value in Horizontal RMSE whereas the
significant better vertical accuracy was achieved of
approximately 14 percent compared to equal-weight
model. In addition, their standard deviations of Figure 3: W-ratio
baseline component is comparable as shown in Table2
6. Conclusion remarks
Number H RMSE (cm.) V RMSE (cm.)
Epoch MQ EQW MQ EQW In this paper, the MINQUE procedure has been
14 1.7 1.8 3.3 3.6 tested for VRS technique. The finding show that the
15 1.7 1.7 3.0 3.5 MINQUE procedure can improve the positioning
16 1.6 1.6 2.9 3.5 accuracy and reliability of ambiguity resolution.
17 1.6 1.6 3.0 3.7 Further tests should be carried out with some other
18 1.6 1.6 3.0 3.7 datasets to confirm the conclusion.
19 1.6 1.7 3.1 3.5
20 1.6 1.6 3.1 3.5
References
Table1: RMSE results for each stochastic model.
[1] Charoenkalunyuta, T., Satirapod, C., Li Y. and
.From Figure 2 and 3 it is obvious that the Rizos, C. 2012. An investigation of the effect of
ambiguity validation statistic, both F-ratio [6] and W- Ionospheric models on the performance of Network –
ratio [7] obtained by MINQUE procedure, are based RTK GPS in Thailand. International Journal of
consistently larger than those of the standard Geoinformatics, 8(4), 61-66
procedure. This result suggest that the confidence in [2] Chen, H.Y., Rizos, C. and Han S., 2004. An
the ambiguity resolution is improved. instantaneous ambiguity resolution procedure suitable
for medium-scale GPS reference station networks
Number H SD. (cm.) V SD. (cm.) Survey Review, 37(291), 396-410
Epoch MQ EQW MQ EQW
[3] Al-Shaery, A., Lim, S., Rizos, C., 2011
14 0.1 0.1 0.2 0.1 Investigation of Different Interpolation Models Used
15 0.1 0.1 0.1 0.1 in Network-RTK for the Virtual Reference Station
16 0.1 0.1 0.1 0.1 Technique, Journal of Global Positioning Systems,
17 0.1 0.1 0.1 0.1 Vol. 10, No.2, 136-148
18 0.1 0.1 0.1 0.1 [4] Satirapod, C., Wang, J. and Rizos, C. (2002) A
19 0.1 0.1 0.1 0.1 simplified MINQUE procedure for the estimation of
20 0.1 0.1 0.1 0.1 variance-covariance components of GPS observables.
Survey Review, 35(286), 582-590
Table2: S.D. results for each stochastic model. [5] De Jonge, P. and Tiberius C., 1995 The LAMBDA
method for integer ambiguity estimation:
implementation aspects. Publications of the Delft
Computing Centre, LGR-Series 12 (12), 1-47
[6] Euler, H. J., and Schaffrin, B., 1991 On the
measure for the discernibility between different
ambiguity solutions in static-kinematic GPS mode.
IAG Symposia no.107, Kinematic Systems in
Geodesy, Surveying and Remote Sensing, 285-295
[7] Wang, J. Stewart, P Tsakiri M. 1998 A
discrimination test procedure for ambiguity resolution
on-the-fly Journal of Geodesy 72, 644-653

Figure 2: F-ratio

Potrebbero piacerti anche