Sei sulla pagina 1di 6

Geolocation based on Measurements Reports for

deployed UMTS Wireless Networks


Tiago Silva Pereirinha

Antonio Rodrigues

Pedro Vieira

Instituto Superior Tecnico


University of Lisbon
Lisbon, Portugal
tiago.pereirinha@ist.utl.pt

Instituto de Telecomunicaco es
Instituto Superior Tecnico
University of Lisbon
Lisbon, Portugal
ar@lx.it.pt

Area
Departamental de Engenharia de Electronica
e Telecomunicaco es e de Computadores
Instituto Superior de Engenharia de Lisboa
Lisbon, Portugal
pvieira@deetc.isel.pt

AbstractAlthough geolocation of mobile phones on UMTS


networks is possible, it usually entails high implementation
expenditure. In deployed networks, for purposes like traffic
analysis and network maintenance, drive-tests can be taken,
albeit with associated operational costs as well. In this thesis, an
alternative geolocation method using Measurement Report Messages (MRM) has been studied. Using real measurements from a
single User Equipment in Lisbon, a data-abstract algorithm was
developed, which takes as input MRM and Node B information,
performs indispensable initial calculations, estimates the position
using a non-linear recursive least squares trilateration method,
and outputs the results in several types of data. The results
were tested using the corresponding drive-tests collected GPS
coordinates, and a median positioning error of 272 meters of
was achieved, using 32.3% of the MRMs. Certain parameters
were found to regulate the positioning error-MRM usability
rate tradeoff, and general results were shown to be affected by
various factors, such as multipath degradation, low data availability, non-ideal synchronization, and limited number of cells
available. Nevertheless, results were promising and a way was
paved for further improvement using integrated motion modelbased Kalman filtering. The algorithm was also left theoretically
prepared to receive and analyze LTE MRMs when they eventually
become available. The thesis concludes with an analysis on the
positive impact of the algorithm and an outlook on the future of
geolocation technology implementation on deployed UMTS and
LTE networks.

I. I NTRODUCTION
Mobile communications, since their appearance and first
steps as an industry, have rapidly risen and currently hold a
significant importance at social and financial levels. Together
with the growth of cellular systems and technologies however,
an effort is being made to consistently improve the quality and
quantity of services provided to the end user. Regarding that
theme, one of the problems that has sparked a noteworthy
amount of research and development is the finding of ways to
provide a precise location of the User Equipments (UE). For
mobile operators, UE location is very useful and important, for
reasons such as network design and maintenance, traffic map
generation, handover algorithm improvement and accurate radio propagation modeling, through Key Performance Indicator
(KPI) analysis. For this purpose, drive-tests are taken, in which
UE measurements and geographical data is acquired simultaneously, for posterior analysis. However, drive-tests entail

considerable operational costs. To circumvent this drawback,


Measurement Report Messages (MRMs) generated by the UEs
and reported to the network, have been studied as a way of
geolocating the handset.
The main objective is to study and develop a complete
data-abstract platform, based on the OTDOA (Observed Time
Difference of Arrival) algorithm, to geographically locate User
Equipments on already deployed UMTS networks, using real
MRMs collected in Lisbon. The point is to achieve a solution
that delivers position estimates reliable enough to be used on
network analysis and maintenance, among other applications,
in order to ultimately reduce the quantity of drive-tests being
performed, successfully suppressing a considerable portion of
the associated operational costs. The concept has already been
studied, albeit within different scopes or using different types
and variety of data [1], [2]. Also, all of the solutions are
proprietary, and therefore not directly available for study.
This article is structured as follows: Section II covers the
algorithm structure from input loading, parsing and positioning
to output generation. Section III explains how the resulting
position estimates were assessed in terms of accuracy, followed by an analysis of the results yielded. Finally, Section
IV summarizes the main conclusions from this study and gives
some pointers regarding future work.
II. A LGORITHM S TRUCTURE
A. Input data
The input data necessary to calculate, within an acceptable
error margin, the estimates of the physical position of the
UE is the Node B information and the MRMs. The Node B
repository, henceforth the cell dump, must contain the necessary data, which are the names, identification numbers (Cell
IDs), Primary Scrambling Codes (PSC), Primary Common
Pilot Channel (P-CPICH) powers, geographic coordinates and
antenna heights off all the cells of the network. As for the
MRMs, each one must contain the event that triggered the
report, the Cell IDs of the cells in the active set, the PSCs of
the cells in the active and monitored sets, and for each one of
those radio links, the Ec/N 0, Received Signal Code Power
(RSCP), frame offset (OFF) and chip offset (Tm) parameters
[3].

Although additional data can be used in order to further


optimize the algorithm, said data may not be provided, so only
the utmost essential elements needed to calculate the position
are going to be taken into account at this level.
After all the data is successfully loaded and parsed, the steps
that follow are repeated for each MRM, and each one generates
a set of coordinates for the position of the UE, assuming that
said MRM complies with all of the following requirements of
trilateration methods, which are used twice over the course of
the algorithm:
a) The MRM provides measurements from at least three
different cells.
b) All three (or more, if available) of those cells must be
from different Node Bs (henceforth, sites), i.e., they must
not share the same location.
c) All three (or more, if available) of those cells must be not
be located in a way that makes them co-linear in space.
B. Initial Calculations
Prior to the actual position estimation method, some
calculations and verifications are essential, mainly for
unequivocally identifying sites and cells, eliminating
redundant measurements and calculating Observed Time
Differences (OTDs) and a first estimate for the position.
1) Cell identification: Since of the cells in the monitored
set only the PSC is known, and since PSCs are only ranged
[0, 512] (i.e., the PSC is not enough to uniquely identify the
cell), to determine which cells are in fact communicating with
the UE, it is necessary to ascertain their Cell IDs. So, the
coordinates of the strongest (in terms of Ec/N 0) cell in the
active set are used as reference, and for a PSC n being tested,
all the cells in the cell dump with that PSC are checked, and
the cell with a PSC n which is closest to the reference is
assumed to be the one linked to the UE. This proximity-based
approach is effective provided the network design, in terms
of PSC assignment, is acceptable. The distance between the
reference and the tested cells is calculated using the Haversine
formula.

is important to the convergence of the RLS method, the first


estimate is calculated using a more complex method, via a
modified Okumura-Hata model and geometric trilateration.
The Okumura-Hata model [4] uses empirical equations to
estimate the pathloss, on which one of the variables the
distance, among others. Thus, by applying a different equation,
which uses the cells P-CPICH transmission power T x and the
MRMs reported RSCP to estimate the pathloss P L [5]
P L[dB] = T x[dBm] RSCP[dBm]

and by rearranging the Okumura-Hata equations, the distance di from a cell i to the UE can be calculated by
P L 69.55 26.16 log10 f + 13.82 log10 hB + CLC
44.9 6.55 log10 hB
di = 10
(2)
where f is the transmission frequency, hB is the cells
antenna height, and CLC is the model-specific parameter for
large cities.
Then, using the three strongest cells coordinates,
(x, y)A,B,C , converted to the UTM system [6], and the three
calculated distances dA,B,C , the UTM coordinates for the first
position estimation (x, y)U E can be obtained using geometric
trilateration:

2
2
2

(xU E xA ) + (yU E yA ) dA = 0
2
2
(3)
(xU E xB ) + (yU E yB ) d2B = 0

2
2
2
(xU E xC ) + (yU E yC ) dC = 0
The coordinates are obtained after rearranging the system
of equations and solving it by applying Cramers rule.
4) OTD calculation: Finally, the SFN-SFN Observed Time
Difference [3], in chips, for a connection between the UE and
cell i, reported by the k-th MRM, using the frame offset OFF,
which has a range of [0, 1, . . . , 255], and the chip offset Tm,
which has a range of [0, 1, . . . , 38399], can be calculated with
OT Dk (i) = 38400 OFF(i) + Tm(i)

2) Elimination of redundant measurements and Site


identification: As mentioned before, since the positioning
method is based on trilateration, measurements made by the
UE to cells located in the same place are redundant, as two
equations between the same two locations and the UE are
not linearly independent. Therefore, all links within an MRM
must be tested, and for each redundant link pair is found, the
one with the lowest measured RSCP is eliminated. At this
point, a site repository is also created, to assist the algorithm
in posterior phases.
3) First position estimation: The non-linear iterative RLS
method used for position requires that an initial estimate for
the position is provided. Using the strongest cell position [1] or
the centroid of the triangle formed by the strongest three cells
are possible options, but since the accuracy of this estimate

(1)

(4)

C. Positioning Cycle
Using the original measurements, cell information and additional parameters defined on the sections above, the algorithm
enters an iterative cycle, which produces, as a result, the
final estimated coordinates of the UE position. The cycle is
composed by three main steps:
Estimate propagation delays of the connections between
the cells and the UE (using newly estimated coordinates),
Calculate Relative Time Differences (RTDs), based on
OTDs and Propagation delays,
Run a non-linear RLS algorithm, which generates new
coordinates.
This iterative process is repeated a pre-defined number, K,
of times, or until the coordinates stabilize. Stabilization, in
this context, is assumed when coordinates do not change for

5 iterations of the cycle. The first iteration will use the first
position estimate calculated above, and the ones after will
use the estimates produced by the previous one.
1) Propagation delay and RTD calculation: The first step is
calculating the propagation delays for every link of the MRM.
In the k-th MRM, the propagation delay k (i), between the
UE and a cell i, is given by
1 q
(5)
k (i) = (xc(i) xue )2 + (yc(i) yue )2
c
where c is the speed of light, and (x, y)c(i) and (x, y)ue
are the coordinates of the positions of cell i and the UE,
respectively.
Then, using the OTDs and the propagation delays, RTDs
are calculated. On an MRM k, and RTD sample, between two
cells belonging respectively to sites i and j, can be obtained
as follows:
RT Dk (i, j) = OT Dk (i) OT Dk (j) (k (j) k (i)) (6)
RTDs, like OTDs, are in chips, which makes the conversion
of k from seconds to chips necessary. In UMTS, the chip rate
is 3.84 Mchips/s. Ergo, one chip in seconds corresponds to
the inverse of that figure, which is 0.26 s. Also, it has to be
taken into account that RTDs have a range of [0, 1, . . . , 256
38400 1]. So, incorporating those notions, the expression for
RTD calculation becomes

and
c(i, j) = (d(i, j) + 128 38400) mod (256 38400)
128 38400

(10)

g
RTD(i,
j) is the median of RTD(i, j), and p is the
exponent that defines the ratio of outliers eliminated. The dot
before the parameters denotes element-wise exponentiation.
RT Dk0 (i, j) are calculated for every cell pair involved in
the usable measurements of the MRM.
2) Non-linear RLS trilateration: The final step takes all
the parameters above and uses them to perform a trilateration,
producing the coordinates for the physical position of the UE.
Said trilateration is executed as a non-linear RLS optimization
problem, which follows an algorithm called Trust-RegionReflective (TRR). This algorithm is a subspace trust-region
method and is based on the interior-reflective Newton method
described in [7], [8]. Each iteration involves the approximate solution of a large linear system using the method of
preconditioned conjugate gradients. A system of equations
is constructed, and takes an initial estimate of the position
then performs a number of iterations, each time applying a
correction to the previous value, reducing the sum of squared
residuals of the different equations:
min kf (x, y)k22 = min f1 (x, y)2 +f2 (x, y)2 +. . .+fn (x, y)2
x,y

x,y

(11)
f
(x,
y)
is
a
system
of
n
equations,
ranged
from
f
(x,
y)
1
k (j, i) 
mod
(25638400)
RT Dk (i, j) = OT Dk (i, j)
to
f
(x,
y).
The
common
unknown
to
all
the
equations,
is
the
n
0.26 10 6
(7) position of the UE, (x, y)ue . Each equation uses measurements
Each MRM k analyzed measuring sites i and j will produce from an MRM k reporting data from two cells/sites i and j:
an RTD sample, RT Dk (i, j). These samples are saved in a
fk,i,j (x, y) = rk,i,j dk,i,j (x, y)
(12)
three-dimensional matrix called RTD array. The first two
dimensions of the matrix are the sites. Each position of the
where
matrix is an array containing all the samples already calculated
for the two sites. The matrix, conceptually, is strictly upper


rk,i,j = c 78 OT D(i) OT D(j) RT Dk0 (i, j) (13)
triangular.
The RTD values used on the next step will not be the
and
samples calculated in this step for this k-th MRM, but actually
a modified median of all the samples already calculated,
q
in MRMs [1, . . . , k], between the two sites in question. A
dk,i,j (x, y) = (xc(i) xue )2 + (yc(i) yue )2
q
(14)
simple median is not suitable because there might be error(xc(j) xue )2 + (yc(j) yue )2
contamined measurements amongst the data. The solution is
to attenuate any outliers that might appear on the distribution
So, an MRM k containing n valid measurements from
of values [1]. So, instead of calculating the median of the array
the UE to cells/sites (1, 2, . . . , n), ranked in order of signal
0
RTD(i, j), the final value RT Dk (i, j) between sites i and j
strength, will produce (n 1) equations:
is obtained as follows:

fk,1,2 (x, y)

fk,1,3 (x, y)
g
RT Dk0 (i, j) = sgn c(i, j) |c(i, j)|.p + RTD(i,
j)

fk (x, y) =
(15)
(8)

..

mod(256 38400)
.
fk,1,n (x, y)

where
g
d(i, j) = RTD(i, j) RTD(i,
j)

(9)

Naturally, if the MRM reports 4 or more valid measurements, the system is overdetermined. The reason why only

(n 1) equations are generated from n measurements is to


attenuate multipath propagation effects - only RTDs involving
the cell with strongest reported RSCP are used, even though
RTDs for every site pair on the MRM are calculated.
Besides a new position estimation, supplementary information data is produced: the squared norm of the residual, RN ,

RN =

n
X

f1 (x, y)2 + f2 (x, y)2 + . . . + fi (x, y)2

(16)

i=1

and the exit flag of the RLS/TRR algorithm.

the raw data of the MRMs, including all values and


structures used to produce the positions estimates,
a CSV table, in which each row corresponds to a measurement, and contains the latitude and longitude coordinates
of the UE position, the active set size, and the RSCP and
Ec/N 0 of the strongest measurements,
two Google Earth KML files, for visual reference of the
data, containing all the points yielded by the algorithm.
In one file, the pins are color-coded according to the size
of the active set, and on the other, according to the their
Ec/N 0 values.
III. A SSESSMENT AND R ESULTS

D. Validity Criteria
After the end of the cycle, the algorithm is in possession of
the final estimate for the UE position. However, not all of the
positions will be valid, since for a given MRM the RLS/TRR
algorithm might not converge to a stable mathematical solution, or the measurement data might be corrupted. So, the
algorithm will attribute a flag of validity to the MRM, branding
it valid or not valid according to the following verifications:
A negative RLS/TRR exit flag means that the the method
was unsuccessful. In that case, it can be assumed that the
resulting position estimation is compromised, and thus is
not valid.
If the positioning cycle reaches K iterations without
stabilizing, it means that the cycle has been consistently
producing a different position for every iteration. Consequently, it has to be assumed that the position, whichever
it is, might not be valid.
Each MRM k analyzed measuring sites i and j will
produce an RTD sample, RT Dk (i, j), which is then
saved in the three-dimensional matrix RTD array. Then,
an RTD value, RT Dk0 (i, j) is extracted from that array
for use in the RLS/TRR algorithm. RTD values between
the same sites, ideally, should be the equal for every
MRM, regardless of other factors. However, there might
be MRMs that have error-contaminated values. So, the
final validity verification is testing how much the kth MRM RTD samples diverge from the RT D0 values.
Given that difference RT D ,
RT D = |RT Dk (i, j) RT Dk0 (i, j)|

(17)

the UE position produced for that MRM is considered


invalid if the following inequality is true, for a pre-defined
maximum error tolerance value RT D :
RT D RT D

(18)

E. Output
The very final phase of the algorithm is generating output.
Besides the UE position coordinates, the algorithm provides
other information that can be used for post-processing analysis
of the estimates and of the performance of the algorithm itself.
The outputs generated are:
text logs of every operation performed,

A. Assessment method
To test the accuracy of the estimates, they were tested
against GPS measurements acquired as the MRMs were
recorded, i.e., during a drive-test. For every MRM which
yielded a location, the algorithm uses its timestamp to find
the accompanying GPS position for comparison. However,
due to delays and clock drifts, there might not be an exact
match between the MRM timestamp and a GPS one. Thus, a
weighted linear interpolation is used to find the best value for
the coordinates, and only then is the positioning error between
the estimates and the real positions, using the Haversine
formula.
After all the MRMs have gone through this comparison,
many important indicators can be calculated and generated,
including error median values and cumulative distribution
function graphs.
B. Results analysis
Only MRMs with three or more usable measurements were
considered, since it is not possible to perform trilateration with
only one or two beacons. Table I explicits how many MRMs
were analyzed and, of those, how many produced positions
considered valid as per the criteria of Section II-D. It can
be easily observed that the fact that only three or more cells
MRMs are usable quickly eliminates a very large number
of MRMs, which is a significant weakness of any OTDOAbased algorithm. From the remaining MRMs, even though all
of them produce position estimates, an additional amount are
rendered invalid by the validity criteria, but it is important to
note that there is an implied trade-off between the number of
valid estimates and the widening or narrowing of the validity
criteria - narrower criteria yield less estimates, but, in theory,
closer to the actual position of the UE.
Figure 1 shows the Cumulative Distribution Function of the
positioning error. The error correction median obtained was
272.62 meters, which falls within the accuracy ranges for a
OTDOA solution in urban environments, [50 300] meters
(specific range for urban environments) [9]. It can also be
seen that, under the same conditions (no RTD values known
a priori and no implemented Kalman Filter), the algorithm
performs similarly to that of [1].
Even though the results are acceptable for the input data
provided (only cell and MRM data as opposed to cell, MRM,

TABLE I
MRM ANALYZABILITY AND VALIDITY CRITERIA - COMPLIANT VALUES

AND PERCENTAGES

Total MRMs
MRMs w/ 2 measurements
MRMs w/ 3 measurements
Analyzable MRMs [%]
Valid estimate MRMs
Invalid estimate
Valid MRMs [%]

1123
630
493
43.900
159
334
32.252

The method used in the assessment, might introduce


errors when generating the error measurements.
OTD (and, consequently, RTDs) have resolutions of one
chip, or rather, 78 meters, which is a considerable distance.
Since MRMs are based on handover events, due to
their nature there will be bursts of MRMs reporting the
same cells. If those measurements all generate wrong
positions, there will be a big impact on general algorithm
performance.

The final algorithm generated all outputs successfully. Figure 2 provides a screenshot of Google Earth loaded with one
of the output KML files.

Fig. 1.

Cumulative Distribution Function of the positioning error

U-RNTI, RTT and elevation data), the fact that it does not yield
positioning estimates with even smaller positioning errors can
be attributed to a number of factors:

The data was collected in Lisbon, which is a city with


irregular terrain, with little to no spatial planning and
heterogeneous architecture, all traits which exacerbate the
effects of multipath.
OTDOA-type algorithms perform worse on urban environments [9].
Due to trilateration performance requirements, two-orless-cell-MRMs are of no use in this context, which
renders more than half of the reports useless straight from
the onset.
Finding the cells by PSC proximity analysis is not infallible - there is always the possibility of the closest cell
with a given PSC not being the cell on the measurement
with that PSC. That probability grows proportionally to
the density of cells in the area.
All calculations throughout the algorithm are performed
on 2D, as no elevation data was provided.
Many calculations are made on the course of the algorithm, most of which involve approximations.

Fig. 2.
file.

Screenshot of Google Earth UE running the Ec/N 0 KML output

IV. C ONCLUSION
The developed algorithm proves to be a very promising
starting point to a fully usable post-deployment low-cost
location solution, leaving room for improvement in order to
ultimately achieve a tool that eliminates almost completely
the need to perform most drive-tests. Key additions to be
considered include:

Using elevation data would be useful to modify the


algorithm to work in 3D.
The development, and subsequent integration, of a motion
model based-Kalman filter, which would be used to
further improve the estimated positions by using a locally
linear prediction model of successive positions [10], [11],
to provide a better control over non-convergent position
estimates generated by the algorithm.
Given a powerful enough machine, the algorithm can be
easily modified to run various traces at once, and in realtime.
If LTE MRMs are provided, they can be used to study
modifications to the algorithm in order to make it LTEcompatible.

R EFERENCES
[1] C. Brunner and D. Flore, Generation of Pathloss and Interference Maps
as SON Enabler in Deployed UMTS Networks, VTC Spring 2009 IEEE
69th Vehicular Technology Conference, pp. 15, 2009.

[2] C. Ubeda,
J. Romero, and J. Ramiro, Evaluation of a time-delay based
geolocation algorithm in real UMTS networks, 2010 Fifth International
Conference on Broadband and Biomedical Communications, pp. 14,
2010.
[3] 3rd Generation Partnership Project, Universal Mobile Telecommunications System (UMTS); Physical layer; Measurements (FDD) (3GPP TS
25.215 version 11.0.0 Release 11), reference RTS/TSGR-0125215vb00,
Tech. Rep., 2012.
[4] A. Molisch, Wireless Communications, Second ed. John Wiley & Sons,
2011.
[5] 3rd Generation Partnership Project, Universal Mobile Telecommunications System (UMTS); Radio Resource Control (RRC) (3GPP TS 25.331
version 11.6.0 Release 11), reference RTS/TSGR-0225331vb60, Tech.
Rep., 2013.
[6] S. Dutch, Converting UTM to Latitude and Longitude (Or Vice
Versa), Webpage, 2012. [Online]. Available: http://www.uwgb.edu/
dutchs/usefuldata/utmformulas.htm
[7] T. Coleman and Y. Li, An Interior, Trust Region Approach for Nonlinear Minimization Subject to Bounds, SIAM Journal on Optimization,
vol. 6, pp. 418445, 1994.
[8] , On the Convergence of Reflective Newton Methods for LargeScale Nonlinear Minimization Subject to Bounds, Mathematical Programming, vol. 67, no. 2, pp. 189224, 1994.
[9] A. Kupper, Location-based Services: Fundamentals and Operation.
John Wiley & Sons, 2005.
[10] M. Hellebrandt and R. Mathar, Location tracking of mobiles in cellular
radio networks, IEEE Transactions on Vehicular Technology, vol. 48,
no. 5, pp. 15581562, 1999.
[11] D. Catrein, M. Hellebrandt, R. Mathar, and M. Serrano, Location
tracking of mobiles: a smart filtering method and its use in practice, in
Vehicular Technology Conference, 2004. VTC 2004-Spring. 2004 IEEE
59th, vol. 5, 2004, pp. 26772681.

Potrebbero piacerti anche