You are on page 1of 6

2004 IEEE lnternatianal Conference on IndustrialTechnology (JCIT)

Location Algorithm for Wireless Sensor Networks


in Industrial Applications
Jagoba Arias, Jes6s Liizaro, Armando Astarloa, Jaime Jimknez, Aitzol Zuloaga
Faculty of Engineering
Department of Electronics and Telecommunications
University of the Basque Country, Spain
Alda. Urquijo s/n, Bilbao
Email: {j tparpejtplaarj,jtpascuajtpjivej,jtpzuiza}j@ bi.ehu.es

Absftact-Wireless sensor networks is an emerging technology,


which has aroused a great interest in the research community in
the last years. The applicability af this technology to industrial environments, which are usually noisy and disturbing, requires that
new algorithm are developped, which are robust enough to overcome the difficulties that appear in these applications. The most
basic sensing activity a wireless sensor network can perform is the
location of its nodes. This basic information can make other tasks,
such as routing, calibration, etc. easier to tackle with, which makes
it a good sensing example for these robust algorithms.
In this paper we describe a distance based location algorithm,
which degrades gracefully as the error made in the estimation of
distances grows. This is the first step to get a group of robust algorithms, which take advantage of the cooperation among sensors to
provide more reliable information.
Index Tems--WireIess sensor networks, location, RSSI

I. INTRODUCTION
In the recent years, the wireless communication industry has
launched new (and cheaper) integrated circuits for providing
connectivity to digital systems. As a result, new applications
based on this technology have appeared, which take advantage
of this new capability. An example of this are wireless sensor networks [ I]. These networks consist of a reIatively high
number of inexpensive (and usually unattended) nodes, which
are able to share data using wireless transceivers. Their weak
nature involves some restrictions to their working conditions:
Being inexpensive nodes, they are prone to failure. They
main force of this network does not lay on the reliability of its node but on their number: being many nodes,
some may fail at a certain moment, but they will restart
afterwards. During their off time some other nodes will
perform their tasks for them.
Having unattended nodes means that there will be power
consumption restrictions for the nodes. Rechargeable batteries using solar cells or some other renewable power supplies are required to maintain the nodes working.
Small and cheap nodes cannot perform complicated computational operations. Nodes in these networks will have
little memory and low computing capabilities - which
eases the previously mentioned problem of power consumption.
These restrictions must be overcome to provide some services to the user:
0-7803*8662-0/O4/!20
.OO 02004 IEEE

The nodes, which are scattered in a relatively wide geographical area, may not be able to communicate with each
single node in the network. Therefore, they must cooperate to move the information from the sources to their sinks.
The user must be able to know the value of each measurement as fast as possible.
I The sensors must be. able to calibrate themselves automatically - it would be a nightmare, if a human operator had
to manually calibrate several hundreds of nodes [ 2 ] .
The whole structure of the network must not suffer if some
nodes die for some time.
Wireless sensor networks have been normally used in some
typical environments:
Habitat monitoring: there is a large [and usually inaccessible) area, where some protected species live. The objective
of the network is to control the presence of these animals,
and the different magnitudes that may affect their quality
of life (temperature, humidity, etc) [3].
Area surveillance: in this case, the aim of the network is
to detect the presence of hostile elements in a certain area,
sending alerts to a surveillance center [4].
Medical care: this time the wireless sensors are installed
in each patient of a medical facility. If the sensors observe
some kind of abnormal behavior of the patients constants,
they will send an alarm to the nearest doctor [ 5 ] .
However, these sensor networks may aIso be applied in the
industry to control the behavior of large machines (or even large
groups of cooperating machines). In this case, each node will
have sensors and actuators, so that both kinds of action can be
taken: on one hand, different magnitudes can be measured and,
on the other hand, the node can perform some operations on the
system under control.
Most of the operations described so far require that the location of cach node is known: it is useless to know that sensor 47h
has measured a temperature of 34OC,if we do not know where
this sensor is, routing procedures may be easier to perform if
the geographical structure of the network is known, etc [6]. In
this paper, a location algorithm is described, which may be used
in industrial applications, in order to improve the performance
of wireless sensor networks used in this area.
This article is organized as follows: in section II, the work
related to localization in wireless sensor networks is described.
757

The location aIgorithm is described in section III and its performance under different circumstances is measured and quantified in section N.The final conclusions of this work can be
found in section V.
11. RELATEDWORK

In the last years, different approaches have been followed to


find the geographical position of wireIess nodes within a network. Depending on the kind of information used, these methods can be classified as:
Angle of Arrival (AoA) detection systems: this approach
needs that each beacon has a directive antenna, which can
be (mechanically or electronically) rotated around its axis.
When one of the beacons detects a peak in the received
power, it is assumed that the target is in the direction
pointed by the antenna [ 7 ] , [SI, [SI. This location approach
does provide a reasonably high accuracy but it is not suitable for wireless sensor networks, because it requires moving parts or complex antenna configurations, which make
the nodes expensive.
4 Time of Arrival (ToA) detection systems: these systems
measure the amount of time a wave needs to get from a
beacon to the target. Thus, the distances between the target and a set of known beacons can be used to estimate
the actual position of the target [ 101, [ll]. However, this
h n d of scheme only allows the use of RF signaIs when the
distances are high (thus, the lapse of time of the travelling
signal is noticeable). Indoors, or in places where the presence of obstacles reduces the operative range of the beacons, this approach requires the use of ultrasonic waves,
which are much slower. However, this makes them more
sensitive to ambient noise and two different communication systems must be supported: the ultrasonic ranger and
the RF interface used for synchronization and data sharing
functions.
Power of Arrival (PoA) detection systems: these systems
use the power of the incoming radio signal to estimate the
distance between the target and the beacon. The presence
of obstacles will affect negatively to the measured power,
so the distance estimation will not be accurate [12], [ 131.
Frequency of Arrival (FoA) detection systems: they are
based on the Doppler effect: as the target moves, the signals it transmits suffer from a frequency deviation (called
Doppler shift), which depends on the relative velocity of
the emitter to the receiver, Thus, using several receivers,
the exact position of the target and its velocity can be calculated 1141, [15].
The simplest method described in this section is the Power
of Arrival scheme, because it does not need any kind of additional hardware to perform the location procedure: most receivers have an RSSI (Received Signal Strength Indicator) indicator, which enables the measurement of the incoming power.
However, this is also the most inaccurate approach, due to the
high variability of attenuation in a wireless channel.
All the described methods require a special locating infastructure, i.e. a certain number of nodes, whose exact position
is known, which arc usually known as beacons. These bea-

cons will provide the reference system and the necessary information to find the rest of nodes in the network. In general,
for wireless sensor networks, the most commonly used location methods are based on distance estimation (PoA and ToA),
because the detection of angles or frequencies usually require
more complex equipment (antennae, whose radiation pattern
rotates mechanically or electronically, complex signal processing circuits, etc). Among the other algorithms that calculate the
position of a node using distance estimations, the most typical
ones are described in the following subsections.

A. The Random Algorithm


This technique is the simplest approach we can use to calculate the position of a node, if we know its distance to some
beacons. It consists in establishing a region of interest, which
is the smallest box that contains all the beacons. The estimated
postion of the target is chosen using a uniform random variable.
This, of course, is not a very refined method for finding the position when the distance estimation is accurate, but it gives an
idea of how good a location algorithm is: if the result is poorer
than that given by the random algorithm, it means that it is not
worth implementing - the results do not justify the computational cost.
E. Triangulation
This algorithm is best understood in a 2D space. The beacons are divided into groups of three nodes. Now, for each
group we can draw three circumferences, whose centers are at
the beacons positions and their radii coincide with the estimated distances to the target. In the ideal situation, where the
distances have been measured with no error, the three circumferences will have a singIe common point. However, this will
not be the realistic case. To solve the problem, two of the three
circumferences are taken and their two intersection points are
found. Now, the third beacon is used to decide which point is
closer to this third circumference. After performing the same
operation with all the groups, the final solution is obtained calculating the average point of all of them.
The tridimensional generalization requires groups of four
beacons instead of three and spheres instead of circumferences.

C. Least Squares Algorithm (KQ)


This algorithm uses an LSQ approximation to obtain the optimum solution, which minimizes the location error. To do so,
it simplifies the expression of distance - which is a non linear
equation. The minimization problem consists in N equations,
where the estimated distances (d,) and the positions of the beacons ((q,
yi)) are related to the position of the target using the
following relationship:
d:

(.-.a)

2
3:

2.

f (y - y i )

+ xi - 2x21 + yz + y;
2

- yyi

(1)

If we subtract the last equation from the rest, the x2 and y2


terms disappear, and the result is a linear system. Therefore, a
simple LSQ estimation can be made to find the best solution for
the system. This scheme can be used in a 3D space.
750

to K means that the last movement has been too brusque.


Therefore, the conversion parameter y must be updated,
using a certain learning rate A:

Now, the position of the target is updated, using the convertion parameter X E ( 0 , l ) :

I
I

+-F?

I
I

Unknown
Node

Beacon

Fig. I

Physical interpretation of the m


< vector field.

111. ALGORITHM
OVERVIEW

The main force of wireless sensor networks is redundancy: as


many nodes measure the same events, the errors made in these
measurements can be reduced by comparison. In our particular
case, many nodes will try to estimate their distance to the target
(which is in fact our event). The more distance estimation beacon position pairs - we get, the more information we have
to calculate the node position. Now, the key point is how to use
this information. The method described in this paper proposes
a metaphor: we are going to suppose that the target is tied to
the beacons (whose positions 3 are fixed) with some springs.
These springs will be different from each other and they will
have different equilibrium lengths (the estimated distances di).
Now, the target will be assigned an energy, corresponding to
the elastic energy of the springs (see figure I). The optimum
target position is that, which minimizes this elastic energy. To
find this point, the target is set free and the springs will move
to this equilibrium position. For this purpose, the overall elastic
force is calculated from the current position of the target (Y)in
this way:

The expression shown in equation 2 can be expressed using


only two terms: a scalar function (#a), which depends on the estimated distance di and the position of the beacon, and a unary
vector (fi),which depends on the positions of the vector and
the beacon:
N

a=
i= 1

r$i
%=1

. fi

(3)

Now, vector T8 can be used to show in which direction we


should move the target position, to reduce its elastic energy.
The update of the target position is made using a variable converting factor y. The iterative process is performed following
the following procedure:
Calculate the 3 vector, using the beacons positions and
their distance estimations.
s Compare the current force vector with that obtained in
the previous iteration. The angle they form being close

If the displacement is greater that a previously determined


threshold Athres,a new iteration is performed. Otherwise,
the process stops.
In this algorithm, there are three empirical values, which
must be adjusted: the initiat value of 7 ,the learning rate A and
the desired precision
However, due to the nature of the system, some problems
appear, when tryingto find the optimal equilibrium point for
the target:
There may be several equilibrium positions, where the
overall force is zero - which corresponds to local minima
in the targets elastic energy map. This may lead the target
to a place, where the elastic force is zero, but its energy
does not correspond to the minimum.
Depending on the orography of the energy function, the
convergence of the algorithm may be complicated: the
presence of deep valleys requires a sensible force application policy to take the target to the equilibrium.

IV. RESULTS
A. Simulation Environment

Location algorithms depend heavily on the relative position of nodes, errors made in distance estimations, etc. This
is the reason why a testbench, which provides a stable scenario, where the different location algorithms may be tested,
is needed. For this purpose, the user defines the dimensions of
the scenario and the number of beacons that the node to be located (the target) is going to detect. For each simulation run,
the testbench:
Deploys the beacons and the target in the scenario. The
deploying strategy may be defined manually or an automatic uniform random distribution is used. This ensures,
that for each run the beacons and target are redistributed,
providing new relative positions among them.
Makes the distance estimation. To do so, the exact distance
is calculated and then, some kind of error is added, in order
to corrupt the data and evaluate the performance of a location algorithm in an environment with errors. These errors
may be of different natures, attending to the different characteristics of the different distance estimation techniques:
- Additive errors: a random variable is added to the actual distance value. The stochastic properties of the
error, such as probability distribitioa, mean and standard deviation may be specified by the user, allowing
the estimation of the behavior of the algorithm with
different distance estimation techniques.

759

IO

- Multiplicative errors: this time, a random variable is

multiplied by the actual distance value. Thus, a relative error is modeled, instead of an additive noise,
which was the case in the previous error description.
Using the same node distribution and the same distance estimations, the testbench calculates the position of the target using different algorithms. The difference between the

....--.-.--

,__---

8 -

-.1 6
B;
c

4 -

.-+
-...--

,._.I

..,..*-

.*

,.,_-I

_---I

--

4
6
S l i l Dev. of disrauce csim!dlln lml

8
Srd. Dcv. oiilistanm a t i m i o n Iml

(a) Standard deviation of error for variations of y ( O )


in

(b) Computational cost for variations of y().

(d) Computational cost for variations of A.

(c) Standard deviation of error for variations of A.

Fig. 3. Simulation results for variations of y ( O ) and X

algorithm approaches the solution, the displacements become


smaller, and the algorithm requires more iterations to get the
accuracy determined by
However, this does not mean
that values of X close to 1 are the best solution: a large y will
require a low X to keep displacements in a sensible margin and
viceversa.
Last, but no least, we are going to analyze the effect of the
parameter in the algorithm. This parameter determines
when the iterative process is to be stopped. Therefore, Iarge values of
will result in fast simulations with low accuracy
and smalIer values of this parameter will give better solutions,
at the expense of executing a greater number of iterations. Figure 4(a) shows the evolution of the standard deviation of the
error vector obtained. Here, for the largest value of AT^,.^^
the accuracy is reduced when the distance estimation error is
low. The error does not grow any further due to the intrinsic
uncertainty of the data we have: for a certain distance estimation error, the uncertainty in position can only be overcome by
adding new beacons - which has not been done in this experiment. When the intrinsic uncertainty is big enough to mask the
error of an inappropiate election of &hres, the standard deviation of the error vector tends to the same limit, independently
from the value of AThres.

The computational cost of the algorithm, for different values of &hres is shown in figure 4(b). In this figure, it can be
stated that for low values of
the cost is greater than

for larger values of the same parameter. Large values of this


parameter mean that the iterative process will stop before and
less iterations are performed. The main point in the election of
this parameter is the expected distance estimation error: there is
no point in trying to get the solution with a minimum displacement of 1 mm when the expected error in the measurement of
the distance is 30 cm. Here we must meet a trade-off that enables the system to get a good performance with a reasonabiIy
good accuracy.
When the
parameter is too high, a curious effect
appears: the system needs more time to converge when the distance estimation error is low. The answer to this phenomenon
lies in the very same nature of the problem this algorithm is trying to solve: the location we want to find is in the intersection
of at least four spheres. The appearance of error in the distance
measurements blurs the borders of these spheres, which results in valleys of the error field, which are not as deep as in the
case of ideal measurements. This makes oscillations less likely
to appear, which increases the number of iterations needed to
exit the iterative process. For greater values of ArhTesthis does
not happen, because the stopping condition is met before.

v.

CONCLUSIONS AND

FUTURE
WORK

In this article a location algorithm has been described, which


uses measurements corrupted with errors to provide a robust
method for finding nodes in wireless sensor networks. Con761

'"I

(a) Standard deviation of error.

(b) Computational cost.

Fig. 4. Simulation results for variations of

trary to standard locating algorithms, this has proved to degrade


gracefully as the error in distance estimation grows, This algorithm is ideal for the application of the emerging technologies
in wireless sensor networks in industrial environments, where
the presence of noisy machinery is likely to induce errors in the
sensing devices.
The future work in this area will be focused on improving the
computational efficiency of the algorithm, enabling the most
modest microcontrollers with this locating technology. The robustness of this technique shows that it could also be used in
other fields, where complex information is to be extracted from
erroneous data.

REFERENCES
D.Estrin, L. Girod, G. Ponie, and M. Srivastava. Instrumenting the world
with wireless Sensor networks. In Intenwtiunul Conference on Acoustics.
Speech, and Signul Processing (ICASSP), 200 1.
Kamin Whitehouse and David Culler. Calibration as Parameter Estimation in Sensor Networks. In ACM Intemationrrl Wurhhop on Wireless
Sensors and Applications ( WSNAJin conjunction with MobiCOM, 2002.
Alan Mainwaring. Joseph Polastre, Robert Szewczyk, David Culler, and
John Anderson. Wireless Sensor networks for habitat monitoring. In ACM
Internutional Workhop on Wireless Sensor Networks and Applications
(WSNA) in conjuncrion wiih MobiCOM. 2002.
Brent W. West, Paul G . Flikkema, T h o m Sisk, and Koch George W.
Wmless Sensor Networks for Dense Spatio-Temporal Monitoring of the
Environment: A Case for Integrated Circuit, System, and Network Design. In IEEE CAS Workshop on Wireless Communicufions and Networking, 2001.
R. Kostelnik, P. Lacko, M. Perdoch, T. Torok, I. Ziegfer, M. Bielikovii,
and T.KrajEoviE. BlueMediw -WirelessMedical Data Access Appliance.
Journal of Electrical Engineering, pages 31 1-315.2001.
Jinyang Li. John Jannoni, Douglas De Couto, David Karger. and Robert
Morris. A Scalable Location Service for Geographic Ad-Hoc Routing. In
Pmceedings of rhe 6rh AC'M Intemational Conference on Mobile Cumputing und Networking (MobiCom). pages 12G130, August 2000.
R.G. Stwsfield. Statistical Theory of D.F. Fixing. Journul o f f h e IEE,
94(15):762-770, 1947.
Asis Nasipuri and Kai Li. A Directionality based Location Discovery
Scheme for Wireless Sensor Networks. In FirstACM Intemationul Workshop on Wreless Sensor Networks and Applicutions (WSNA '02),2002.
C.D. McGillem and T.S. Ragpaport. A beacon navigation method for
autonomous vehicles. IEEE Transactions on Vehiculur Technohgy,
38(3):132-139, 1989.
N i s m k a B. Priyantha, Anit Chakraborty. and Hari Bdakrishnan. The
cricket location-support system. In Mobile Computing and Networking,
pages 3243,2000.
Andy Ward, Alan Jones, and Andy Hopper. A New Location Technique
for the Active Ofice. IEEE Personnu1 Communications, 4(5):42-47, October 1997.

[I21 P m v i r Bahl and Venkata N. Padmanabhan. RADAR: An In-Building


RF-Based User Location and Tracking System. In INFUCOM, pages
775-784,2000.
[I31 Roy Want, Andy Hopper, Veronica Falcao, and Jon Gibbons. The Active Badge Location System. ACM Tronsuctions on Infonnnlion System,
10:91-102, January 1992.
[I41 NOAA Satellites and Information.
Argas data cotlection system.
http://noaasis.noaa.gov/ARGOS.
[ 151 Jagoba Arias, Jeslis L i m o , Aitzol Zuloaga, and Jaime Jimknez. Doppler
Location AIgorithm for Wireless Sensor Networks. In Inrernarionol Conference on Wireless Networks (ICWN), pages 509-5 14,2004.

You might also like