Delaunay-triangulation based complete coverage in

wireless sensor networks

Chinh T. Vu
Computer Science Department
Georgia State University
Atlanta, Georgia 30303
Email: [email protected]
Yingshu Li
Computer Science Department
Georgia State University
Atlanta, Georgia 30303
Email: [email protected]
AbstractOne of the major issues in a wireless sensor network
(WSN) is how to cover an interested area. In this paper, we
consider the area coverage problem for variable sensing radii
WSN. With variable sensing range, the difculties to cover a
continuous space in the area coverage problem becomes excep-
tionally harder than covering discrete points in the target (or
point) coverage problem. Very few papers have paid effort for the
former problem. Wang and Medidi have solved the area coverage
problem based on Delaunay triangulation structure [2]. However,
due to the boundary effect their proposed algorithms cannot
always provide complete surveillance for the whole network. In
this work, we improve the work in [2] so that the monitored
area can be completely covered. A theorem conrms that our
improved algorithm provides complete coverage for all the cases.
In addition, the simulation further shows that our energy-efcient
algorithm has an obvious improvement on coverage status with
very small compensation of network lifetime.
I. INTRODUCTION
Sensor network is a special kind of ad-hoc networks which
comprises of large number of sensing, communication and
computation capable devices. One of the major and most
important tasks of those tiny devices is to watch over the
environment (e.g., a forest) or set of subjects (e.g., collection
of precious renaissance paintings) and collect environment
parameters and maybe, further monitor the environment. To
achieve this, the network must have surveillance over all the
environment space or all the subjects. This is the objective of
coverage problem. Since the introduction of coverage problem
in WSN [1], researchers have tried to relax requirements to
achieve better and more efcient results (network lifetime,
fault-tolerance).
The other major issue in WSN is to save energy consump-
tion and extend the network lifetime. Many work aims to
balance the energy consumption among sensors as a method
to extend the network lifetime.
In this work, we improve the work in [2] in order to
provide complete coverage for the network while the energy
consumption among sensors are balanced. We also point out
and correct a mistake in this work. We assume the sensing
region model of a sensor as a circle centered at the sensors
location and sensors sensing range being the radius. We
This work is supported in part by the NSF under grants CCF-0545667
and CCF-0844829.
further assume that sensor could smoothly vary its sensing
range to a maximum cutoff range, denoted by MaxR
s
. We
also assume that a sensor and its neighbors have different IDs,
i.e., within 1-hop vicinity, a node has unique ID.
II. RELATED WORK
The coverage problems for sensor networks can be catego-
rized into three broad types [3] - area coverage (where every
point inside an area is to be monitored), target coverage (where
the main objective is to cover a set of discreet targets), and
breach coverage (the goal here is to minimize the numbers
of uncovered targets or the ratio of uncovered portion to the
whole area). Since Wu and Yang in [7] rst introduce the
coverage problem with the sensor network where sensors may
vary their sensing ranges, some work dedicate effort for this
problem. Most of them [8], [4], [5], [6] try to solve the target
coverage problem since it is easy to verify the coverage status
of discreet target. Very few work [2] solve the area coverage
problem. Some work [4], [5], [6], [7] assumes sensors are only
capable of adjusting their sensing radii to the discreet set of
some pre-dened options. Some other work [8], [2] deals with
network where sensors can smoothly adjust their sensing radii
to any range under some upper threshold.
The work in [7] deals with sensors having two or three
levels of sensing ranges to choose. The work in [4] extends to
allow sensors to vary sensing radii in one of P levels. With this
relaxed assumption, the authors propose three algorithms - one
ILP-based centralized, one greedy centralized and one greedy
distributed - with the objective of maximizing the number of
set covers, each is able to cover the whole set of targets. The
work in [6] makes a step further by providing connectivity for
each of those set covers.
The work in [8] proposes a centralized algorithm to cover set
of targets which schedules sensors based on Garg-K onemann
method. The algorithm has approximation of (1+)(1+logm)
for any > 0 where m is the number of targets.
Except the work in [7], most of the work deals with target
coverage problem and does not take energy into consideration.
In [2], Wang and Medidi propose two distributed algorithms
based on Delaunay-triangulation structure. However, even the
authors have proved that the proposed algorithms could pro-
vide complete coverage, the proof still has aw because it
does not take into account the so-called boundary effect. We
will go into detail about this aw in section III. In this work,
we improve one of their algorithms such that the our new
algorithm can completely cover the whole area.
The rest of the paper is organized as follows: in Section
III, we explain the original algorithm introduced in [2], we
then point out the aw in their algorithm and we x it in
our proposed xed algorithm. In the next section, we dedicate
to simulation illustrating the correctness and efciency of our
algorithm. We rst show the energy model and the simulation
conguration, and in the next subsection we show the simu-
lation results and explain the behind-scene meaning of those
results.
III. THE IDT ALGORITHM
A. Original algorithm proposed in [2]
In this section, we will explain some key aspects of energy
balancing heuristic mentioned in [2]. We name this heuristic as
Original Delaunay-Triangulation algorithm (ODT). We name
our heuristic as Improved Delaunay-Triangulation algorithm
(IDT) which we will explain in more detail in subsection III-C.
Firstly, the authors in [2] introduce a completely distributed
algorithm which requires only one hop neighbor information
to construct an approximate Delaunay-triangulation. It would
be necessary to emphasize that the simple algorithm proposed
in [2] only constructs an approximate Delaynay-triangulation,
not exact Delaunay-triangulation. Thus all the Voronoi cells
or Delaunay triangles built by that algorithm are approximate.
The approximate algorithm works as follows. Each sensor
constructs its Voronoi constellation by drawing bisector of
the line connecting it with a neighbor. Depending on the
relative location of itself, the bisector line and the constellation
constructed so far, the constellation will be modied accord-
ingly. The algorithm to construct Voronoi diagram is very
straight forward and intuitive. After having Voronoi diagram,
the Delaunay triangulation can easily be derived by connecting
pair of sensors whose Voronoi cells share a same border.
The next step, each sensor steps through all Delaynay
triangles which have it as a vertex. For each triangle, the sensor
calculates a weighted centroid based on the location of three
vertices (which include that sensor) and their residual energy.
Having the coordinate of this weighted centroid, the sensor
calculates the distance from itself to the weighed centroid.
The sensor radius will be the larger of its currently assigned
sensing radius and that distance.
This way, all the Delaunay triangles are guaranteed to be
covered by its three sensors locating at its three vertices. Since
the whole area is partitioned into a set of Delaunay triangles,
the whole area is covered too. However, there is still an issue
which we will discuss in detail next.
B. The boundary effect issue
For coverage problem, researchers sometimes have to avoid
so-called boundary effect (e.g., [10]) which is caused by the
fact that the coverage status in the bordering region is different
from the center region. Specially, for algorithms which are
Fig. 1. Boundary effect example
based on the geographical locations of sensors, the border
effect becomes even more severe since the sensors closed to
the area border are not covered by neighbors in all directions.
As discussed in section III-A, we have explained the in-
tuition behind ODT. A question arises is what happens for
sensors locating at boundary region whose Delaunay triangles
are not completely surround them, i.e., there exists an angle
of the sensor that does not belong to any Delaunay triangles?
To answer the above question, let us take an example
shown in Fig. 1. The dashed-lined circles represent sensing
region of three sensors with their maximum sensing ranges. It
can easily be seen that these three sensors completely cover
the whole monitored area (monitored area is represented by
the solid-bordered square). The dashed-bordered triangle is a
Delaunay triangle created by three sensors. The solid-lined
circles are sensing regions of them being assigned by ODT.
Apparently, these sensors cannot cover the whole area with the
new assigned sensing radii. The reason is that the Delaunay
triangle forces those sensors to reduce their radii, therefore,
they cannot reach the border. As the result, the region around
the border is un-covered (or breached). Thus, with sensing
range assignment scheme made by ODT algorithm, it is still
possible that there exists a uncovered region. Next, we explain
our improved IDT algorithm which rules out this possibility.
C. IDT algorithm
Before showing our algorithm, it is necessary to mention
the k-NC rule proposed in [11] which we will employ in
our algorithm. Considering the perimeter circle of sensing
region of a sensor, it can be partitioned into some portions,
and each portion is covered by the same set of neighbors.
For each portion, the number of neighbors which cover that
portion is called the perimeter coverage level of that portion.
The perimeter coverage level of a node is dened in [11]
as the minimum perimeter coverage level of all its perimeter
portions. This parameter usually is denoted by k. Please refer
to Denition 5, in [11] for the formal denition of k-perimeter-
coverage. In this work, we only consider the case where k = 1.
Based on that denition, the authors in [11] propose a rule
which allows a sensor to verify if the network is completely
covered or not based on only 1-hop neighbor information.
The rule claims that if all the sensors in the network are 1-
perimeter-covered, then the network is also 1-covered.
For the sake of ease to explain our IDT algorithm, we
introduce some denitions:
Denition 1: (Uncovered portion) A portion of a sensors
sensing perimeter which is not covered by any of its neighbors
is called an uncovered portion.
Denition 2: (Uncovered nodes) With current sensing ra-
dius assignment, a node is called uncovered node if it is not
completely perimeter-covered by its neighbors.
Denition 3: (Useful neighbors) A neighbor is useful to
a node if when being assigned with the maximum sensing
ranges, it can cover the uncovered portion of that node.
The algorithm pseudo-code is shown below:
Algorithm 1 : IDT algorithm run at sensor s
1: Collect l-hop neighbor information
2: Build the local Delaunay triangulation based on some
existing distributed algorithm like [9]
3: foreach Triangle with 3 sensors s, t, u as its vertices do
4: Determine weighted centroid of .
5: m
x
=
E(s)sx+E(t)tx+E(u)ux
E(s)+E(t)+E(u)
6: m
y
=
E(s)s
y
+E(t)t
y
+E(u)u
y
E(s)+E(t)+E(u)
7: s.R = max(s.R, sm) assign sensing range
8: end for
9: sensors at border increase sensing ranges .
10: if s has the highest priority among all the neighbors &&
NOT completely perimeter- covered then
11: Adjust its sensing range to MaxR
s
.
12: Ask all useful neighbors to adjust their sensing range to
MaxR
s
.
13: end if
After constructing Delaunay triangulation as explained in
Subsection III-C, each sensor may belong to several Delaunay
triangles. Consider a sensor s, it will step through all the
triangles it belongs to. For each triangle, it will calculate
a virtual point named weighted centroid. The coordinates
(x and y) of this point are calculated by line 5 and 6 in
the algorithm 1 where E(s), s
x
, s
y
are sensor ss current
residual energy, x coordinate, and y coordinate, respectively.
t and u are other two vertices of the triangle. Given the
weighted centroids coordinates, s changes it sensing range to
the distance from itself to that weighted centroid. Since there
are several triangles having s as a vertex, each will have it
own weighted centroid. The nal sensing range of s will be the
maximum distance to the furthest weighted centroid. This step
to assign sensor sensing range corresponds with the heuristic
for energy balancing in [2]. The work in [2], however, has
ignored the boundary effect. The ODT algorithm may leave
some region uncovered. We take into account the boundary
effect by letting a sensor which is not completely covered and
all its neighbors who could help to cover the uncovered portion
at the boundary to increase sensing ranges to the maximum
value (MaxR
s
). This step works as follows. After being
assigned sensing ranges in previous step, all the sensors check
if they are 1-perimeter-covered or not by using k NC rule
proposed in [11]. For the ones who already are, they are just
waiting for the request from their neighbors if there are any. To
avoid unnecessary increase of sensors sensing ranges at the
same time, we employ the notation priority. The priority could
be decided by the combination of many other parameters. In
this work we dene priority of a node as E
s
, ID > where
E
s
is sensors residual energy and ID is sensors ID. For a
sensor that has not been 1-perimeter-covered, it rstly waits for
decision (satised with sensing range assigned by ODT or has
already increase sensing range to MaxR
s
) from its neighbors
who have higher priorities. When all those higher-priority
neighbors had made their decision, the node use k NC to
verify its perimeter coverage one more time. If it is still not 1-
perimeter-covered, it increases its sensing range to MaxR
s
. It
then re-evaluates its perimeter coverage status (using k NC
rule) to see if it is 1-perimeter-covered. If it is still not, it will
ask all the useful neighbors to also increases sensing ranges
to MaxR
s
.
Next, we will prove that our improvement indeed guarantees
complete coverage for the network.
Theorem 1: If the monitored area can be completely cov-
ered by the network when all the sensors are assigned
maximum sensing range, then IDT ensures that the whole
monitored area is completely covered.
Proof: Prove by contradiction.
Assume that after all the sensors are assigned sensing range
by IDT algorithm, there still exists an uncovered region G.
Theorem 4.1 in [2] has proven that the part of the area which is
covered by Delaunay triangles will be covered by the sensing
ranges assignment scheme of ODT, thus, G is a region that
is not covered by any Delaunay triangle and G share at least
one border with the monitored area.
Clearly, G shares the border with some sensors sensing
region, denoted by s
1
, s
2
, .., s
l
where each s
i
is a neighbor of
some others and each has sensing range of MaxR
s
(because
of line 11, 12 of algorithm 1).
Consider a point p inside G, meaning s
i
p > MaxR
s
for
all 1 i l. According to our assumption that the whole
area could be covered if all the sensors are assigned with the
maximum sensing range, there will exist a sensor t which can
cover p, meaning tp MaxR
s
, t is assigned a sensing
radius smaller than MaxR
s
, and t is a sensing neighbor of at
least one sensor of set s
1
, s
2
, .., s
l
. Clearly, t G, otherwise,
part of G will be covered by a Delaunay triangle. That means
if being assigned sensing range of MathR
s
, t will cover part
of uncovered perimeter portion of at least one of s
1
, s
2
, .., s
l
.
Or in other words, t is a useful neighbor of at least one
of s
1
, s
2
, .., s
l
. According to line 12 of algorithm 1, t will
increase its sensing range to MaxR
s
, thus p is covered by t
which leads to a contradiction.
Thus, the correctness of the IDT algorithm is proven.
IV. SIMULATION
A. Network conguration, energy consumption model
For energy consumption, we employ the quadratic energy
model as in [12]. That is, when a sensor is in active mode,
the amount of energy consumption for 1 unit of time is
proportional to square of the sensors sensing range. The
formal formula could be as follows:
energy comsumption =
1

R
2
s
where :
is a energy parameter which depends on the character-
istic of a sensor.
R
s
is the sensors sensing range.
The simulation conguration is listed in the following table:
TABLE I
SIMULATION CONFIGURATION
Parameter Value
Network size 800
m
400
m
Number of sensors 100 500
Energy range 100 120 mJoules
Maximum sensing range 100
m
8000
B. Simulation Result
We fairly run the simulation for two algorithms under the
same network (i.e., same number of sensors, sensors locations
and energy assignment and all other simulation parameters are
also the same). For each size of set of sensors, we execute the
simulation 50 times and what is shown in below plots are the
average result from those 50 simulation runs.
Fig 2 presents the comparison on average area coverage
level which is measured by the average of coverage level of
all the points inside the monitored area. For each point, the
coverage level could be 0 (meaning that point is not covered
by any sensor) or 1 (meaning that point is covered by, maybe
more than 1 sensor). The plot shows the percentage of covered
area out of the whole monitored area. As it can be observed,
IDT completely covers the whole area for all time while ODT
cannot. Moreover, the result of ODT is not stable, e.g., the
network with 100 sensors provide better coverage than the
one with 130 nodes. This shows that IDT is more stable than
ODT.
Fig 3 measures the success rate of the two algorithms. If
the result of an algorithm can completely cover the monitored
area, it is a successful execution. In other words, this plot
shows the percentage of successful times of ODT and IDT.
As can be seen, our IDT provides complete coverage for all
cases while ODT does not. Particularly, when the number of
100 150 200 250 300 350 400 450 500
99.75
99.8
99.85
99.9
99.95
100
100.05
Number of sensors
A
r
e
a

c
o
v
e
r
a
g
e

p
e
r
c
e
n
t
a
g
e
ODT
IDT
Fig. 2. Average area coverage level
100 150 200 250 300 350 400 450 500
0
10
20
30
40
50
60
70
80
90
100
Number of Sensors
S
u
c
c
e
s
s

r
a
t
e

(
%
)


ODT
IDT
Fig. 3. Success rate (%)
nodes is 100, the chance for ODT to be successful is only 10%.
This effect can easily be understood considering the boundary
effect explained in Subsection III-B.
Next, we compare the network lifetime affected by the two
algorithms. We rst formally dene the term network lifetime
that is utilized in this work.
Denition 4: (Network Lifetime) The network lifetime is
the duration in which the network could provide the same
coverage status is maintained from the very initial beginning.
Even the above network lifetime denition is not fair for
our IDT since usually ODT cannot provide complete coverage
while IDT can, as shown in Fig. 4 the IDTs network lifetime is
almost the same as that of ODT. Sometimes, the IDTs network
lifetime is even longer. The reason is the boundary effect. That
is, the upper bound of the network lifetime is limited by the
lifetime of sensors close to the boundary regions since the
coverage level of this area is signicantly smaller than that in
the central regions. Thus, if a critical sensor, the one covering
100 150 200 250 300 350 400 450 500
80.5
81
81.5
82
82.5
83
83.5
84
84.5
85
Number of Sensors
N
e
t
w
o
r
k

l
i
f
e
t
i
m
e
ODT
IDT
Fig. 4. Network lifetime
the portion which is not covered by any other sensors, dies,
the network dies as consequence.
It would be seen that if the sensors density at the boundary
area is the same as that in the center area, the performance of
IDT and ODT would be the same. However, it would be hard
to achieve that since sensors are randomly deployed. One way
to reduce the difference between the sensors density in the
boundary area and that in the center area is to deploy more
sensors across the monitored area, i.e., near to the border
and may be outside the border, the performance of ODT must
improve. Apparently, the performance of IDT also enhances.
We believe, even in that case, IDT will still outperform ODT
at least in term of network lifetime. The reason is because of
the fact that the sensor density in the center area is still higher
than that in the boundary area, thus after those sensors who
close to the borders die, the network created by ODT will die
consequently but network created by IDT may still be alive
when the others still may be able to provide 100%-percent
coverage. However, a numerical data of the improvement of
IDT over ODT in this case will be an objective of our future
simulation.
V. CONCLUSION
In this work, we deal with the area coverage problem
with variable sensing radii in WSN by improving the energy
balancing heuristic proposed in [2]. The theorem guarantees
the correctness of our algorithm. The extensive simulation
results prove the efciency and the stableness of our algorithm
in compared with ODT algorithm.
REFERENCES
[1] S. Slijepcevic, M. Potkonjak, Power efcient organization of wireless
sensor networks, IEEE International Conference on Communications ICC,
2001
[2] J. Wang, S. Medidi, Energy Efcient Coverage with Variable Sensing
Radii in Wireless Sensor Networks, Wireless and Mobile Computing,
Networking and Communications, 2007 - WiMOB 2007, White Plains,
NY, USA.
[3] M. Cardei and J. Wu, Energy-Efcient Coverage Problems in Wireless Ad
Hoc Sensor Networks, Journal of Computer Communications on Sensor
Networks, 2004
[4] M. Cardei, J. Wu, N. Lu, M.O. Pervaiz, Maximum Network Lifetime with
Adjustable Range, IEEE Intl. Conf. on Wireless and Mobile Computing,
Networking and Communications (WiMob05), Aug. 2005.
[5] M. Cardei, J. Wu, M. Lu, Improving network lifetime using sensors
with adjustable sensing ranges, International Journal of Sensor Networks,
2006.
[6] S. Yang, F. Dai, M. Cardei, J. Wu, F. Patterson, On Connected Multiple
Point Coverage in Wireless Sensor Networks, International Journal of
Wireless Information Networks, 2006 - Springer
[7] J. Wu and S. Yang, Coverage and Connectivity in Sensor Networks
with Adjustable Ranges, International Workshop on Mobile and Wireless
Networking (MWN), Aug. 2004.
[8] A. Dhawan, C. T. Vu, A. Zelikovsky, Y. Li, and S. K. Prasad, Maximum
Lifetime of Sensor Networks with Adjustable Sensing Range, 2nd ACIS
International Workshop on Self-assembling Wireless Networks (SAWN
2006), Las Vegas, Nevada, USA, June 19-20, 2006.
[9] X.Y. Li, G. Calinescu, P.J. Wan, Y. Wang, Localized Delaunay Triangula-
tion with Application in Ad Hoc Wireless Networks, IEEE Transactions
On Parallel And Distributed Systems, 2003
[10] R. Zheng, G. He, and X. Liu, Location-free coverage maintenance
in wireless sensor networks, Tech. Rep. UH-CS-05-15, Department of
Computer Sience, University of Houston, July 2005.
[11] C.F. Huang, Y.C. Tseng, The Coverage Problem in a Wireless Sen-
sor Network, pp 519-528, Mobile Networks and Applications, 2005 -
Springer.
[12] C. T. Vu, S. Gao, W. P. Deshmukh, and Y. Li, Distributed Energy-
Efcient Scheduling Approach for k-Coverage in Wireless Sensor Net-
works, 25th Military Communications Conference 2006 (MILCOM
2006), Washington DC, USA, October 23-25, 2006.