An Energy Efficient Hierarchical Clustering
Algorithm for Wireless Sensor Networks
Seema Bandyopadhyay and Edward J. Coyle
School of Electrical and Computer Engineering
Purdue University
West Lafayette, IN, USA
{seema, coyle}@ecn.purdue.edu
Abstract— A wireless network consisting of a large number of
small sensors with low-power transceivers can be an effective tool
for gathering data in a variety of environments. The data
collected by each sensor is communicated through the network to
a single processing center that uses all reported data to determine
characteristics of the environment or detect an event. The
communication or message passing process must be designed to
conserve the limited energy resources of the sensors. Clustering
sensors into groups, so that sensors communicate information
only to clusterheads and then the clusterheads communicate the
aggregated information to the processing center, may save
energy. In this paper, we propose a distributed, randomized
clustering algorithm to organize the sensors in a wireless sensor
network into clusters. We then extend this algorithm to generate
a hierarchy of clusterheads and observe that the energy savings
increase with the number of levels in the hierarchy. Results in
stochastic geometry are used to derive solutions for the values of
parameters of our algorithm that minimize the total energy spent
in the network when all sensors report data through the
clusterheads to the processing center.
Keywords- Sensor Networks; Clustering Methods; Voronoi
Tessellations; Algorithms.
I. INTRODUCTION
Recent advances in wireless communications and
microelectro-mechanical systems have motivated the
development of extremely small, low-cost sensors that possess
sensing, signal processing and wireless communication
capabilities. These sensors can be deployed at a cost much
lower than traditional wired sensor systems. The Smart Dust
Project at University of California, Berkeley [14, 15, 16] and
WINS Project at UCLA [1, 17], are two of the research projects
attempting to build such low-cost and extremely small
(approximately 1 cubic millimeter) sensors. An ad-hoc wireless
network of large numbers of such inexpensive but less reliable
and accurate sensors can be used in a wide variety of
commercial and military applications. These include target
tracking, security, environmental monitoring, system control,
etc.
To keep the cost and size of these sensors small, they are
equipped with small batteries that can store at most 1 Joule
[12]. This puts significant constraints on the power available
for communications, thus limiting both the transmission range
and the data rate. A sensor in such a network can therefore
communicate directly only with other sensors that are within a
small distance. To enable communication between sensors not
within each other’s communication range, the sensors form a
multi-hop communication network.
Sensors in these multi-hop networks detect events and then
communicate the collected information to a central location
where parameters characterizing these events are estimated.
The cost of transmitting a bit is higher than a computation [1]
and hence it may be advantageous to organize the sensors into
clusters. In the clustered environment, the data gathered by the
sensors is communicated to the data processing center through
a hierarchy of clusterheads. The processing center determines
the final estimates of the parameters in question using the
information communicated by the clusterheads. The data
processing center can be a specialized device or just one of
these sensors itself. Since the sensors are now communicating
data over smaller distances in the clustered environment, the
energy spent in the network will be much lower than the energy
spent when every sensor communicates directly to the
information processing center.
Many clustering algorithms in various contexts have been
proposed [2-7, 23-28]. These algorithms are mostly heuristic in
nature and aim at generating the minimum number of clusters
such that any node in any cluster is at most
d hops away from
the clusterhead. Most of these algorithms have a time
complexity of
)(nO , where n is the total number of nodes.
Many of them also demand time synchronization among the
nodes, which makes them suitable only for networks with a
small number of sensors.
The Max-Min d-Cluster Algorithm [5] generates d-hop
clusters with a run-time of )(
dO rounds. But this algorithm
does not ensure that the energy used in communicating
information to the information center is minimized. The
clustering algorithm proposed in [7] aims at maximizing the
network lifetime, but it assumes that each node is aware of the
whole network topology, which is usually impossible for
wireless sensor networks which have a large number of nodes.
Many of these clustering algorithms [23, 26, 27, 28] are
specifically designed with an objective of generating stable
clusters in environments with mobile nodes. But in a typical
wireless sensor network, the sensors’ locations are fixed and
0-7803-7753-2/03/$17.00 (C) 2003 IEEE IEEE INFOCOM 2003
the instability of clusters due to mobility of sensors is not an
issue.
For wireless sensor networks with a large number of
energy-constrained sensors, it is very important to design a fast
algorithm to organize sensors in clusters to minimize the
energy used to communicate information from all nodes to the
processing center. In this paper, we propose a fast, randomized,
distributed algorithm for organizing the sensors in a wireless
sensor network in a hierarchy of clusters with an objective of
minimizing the energy spent in communicating the information
to the information processing center. We have used results in
stochastic geometry to derive values of parameters for the
algorithm that minimize the energy spent in the network of
sensors.
II. RELATED WORK
Various issues in the design of wireless sensor networks
−
design of low-power signal processing architectures, low-
power sensing interfaces, energy efficient wireless media
access control and routing protocols [3, 6, 20], low-power
security protocols and key management architectures [29-30],
localization systems [21, 22], etc.
− have been areas of
extensive research in recent years. Gupta and Kumar have
analyzed the capacity of wireless ad hoc networks [18] and
derived the critical power at which a node in a wireless ad hoc
network should communicate to form a connected network
with probability one [19].
Many clustering algorithms in various contexts have also
been proposed in the past [2-7, 23-28], but to our knowledge,
none of these algorithms aim at minimizing the energy spent in
the system. Most of these algorithms are heuristic in nature and
their aim is to generate the minimum number of clusters such
that a node in any cluster is at the most
d hops away from the
clusterhead. In our context, generating the minimum number of
clusters might not ensure minimum energy usage.
In the Linked Cluster Algorithm [2], a node becomes the
clusterhead if it has the highest identity among all nodes within
one hop of itself or among all nodes within one hop of one of
its neighbors. This algorithm was improved by the LCA2
algorithm [8], which generates a smaller number of clusters.
The LCA2 algorithm elects as a clusterhead the node with the
lowest id among all nodes that are neither a clusterhead nor are
within 1-hop of the already chosen clusterheads. The algorithm
proposed in [9], chooses the node with highest degree among
its 1–hop neighbors as a clusterhead.
In [4], the authors propose a distributed algorithm that is
similar to the LCA2 algorithm. In [28], the authors propose two
load balancing heuristics for mobile ad hoc networks. The first
heuristic, when applied to a node-id based clustering algorithm
like LCA or LCA2, leads to longer, low-variance clusterhead
duration. The other heuristic is for degree-based clustering
algorithms. Degree-based algorithms, in conjunction with the
proposed load balancing heuristic, produce longer clusterhead
duration.
The Weighted Clustering Algorithm (WCA) elects a node
as a clusterhead based on the number of neighbors,
transmission power, battery-life and mobility rate of the node
[27]. The algorithm also restricts the number of nodes in a
cluster so that the performance of the MAC protocol is not
degraded.
The Distributed Clustering Algorithm (DCA) uses weights
associated with nodes to elect clusterheads [25]. These weights
are generic and can be defined based on the application. It
elects the node that has the highest weight among its 1-hop
neighbors as the clusterhead. The DCA algorithm is suitable for
networks in which nodes are static or moving at a very low
speed. The Distributed and Mobility-Adaptive Clustering
Algorithm (DMAC) modifies the DCA algorithm to allow node
mobility during or after the cluster set-up phase [26].
All of the above algorithms generate 1-hop clusters, require
synchronized clocks and have a complexity of )(
nO . This
makes them suitable only for networks with a small number of
nodes.
The Max-Min d-cluster Algorithm proposed in [5]
generates d-hop clusters with a run-time of )(
dO rounds. This
algorithm achieves better load balancing among the
clusterheads, generates fewer clusters [5] than the LCA and
LCA2 algorithms and does not need clock synchronization.
In [7], the authors have proposed a clustering algorithm that
aims at maximizing the lifetime of the network by determining
optimal cluster size and optimal assignment of nodes to
clusterheads. They assume that the number of clusterheads and
the location of the clusterheads are known a priori, which is not
possible in all scenarios. Moreover the algorithm requires each
node to know the complete topology of the network, which is
generally not possible in the context of large sensor networks.
McDonald
et al. have proposed a distributed clustering
algorithm for mobile ad hoc networks that ensures that the
probability of mutual reachability between any two nodes in a
cluster is bounded over time [23].
Heinzelman
et al. have proposed a distributed algorithm for
microsensor networks in which the sensors elect themselves as
clusterheads with some probability and broadcast their
decisions [6]. The remaining sensors join the cluster of the
clusterhead that requires minimum communication energy.
This algorithm allows only 1-hop clusters to be formed, which
might lead to a large number of clusters. They have provided
simulation results showing how the energy spent in the system
changes with the number of clusters formed and have observed
that, for a given density of nodes, there is a number of clusters
that minimizes the energy spent. But they have not discussed
how to compute this optimal number of clusterheads. The
algorithm is run periodically, and the probability of becoming a
clusterhead for each period is chosen to ensure that every node
becomes a clusterhead at least once within
P/1 rounds, where
P
is the desired percentage of clusterheads. This ensures that
none of the sensors are overloaded because of the added
responsibility of being a clusterhead.
In [11], the authors have considered a 2-level hierarchical
telecommunication network in which the nodes at each level
are distributed according to two independent homogeneous
Poisson point processes and the nodes of one level are
connected to the closest node of the next higher level. They
0-7803-7753-2/03/$17.00 (C) 2003 IEEE IEEE INFOCOM 2003
have then studied the moments and tail of the distributions of
characteristics like the number of lower level nodes connected
to a particular higher level node and the total length of
segments connecting the lower level nodes to the higher level
node in the hierarchy. We use the results of this paper to obtain
the optimal parameters for our algorithm.
Baccelli and Zuyev have extended the above study to
hierarchical telecommunication networks with more than two
levels in [13]. They have considered a network of subscribers
at the lowest level connected to concentration points at the
highest level, directly or indirectly through distribution points.
The subscribers, distribution points and the concentrators form
the three levels in the hierarchy and are distributed according to
independent homogeneous Poisson processes. Assuming that a
node is connected to the closest node of the next higher level,
they have used point processes and stochastic geometry to
determine the average cost of connecting nodes in the network
as a function of the intensity of the Poisson processes
governing the distribution of nodes at various levels in the
network. They have then derived the intensity of the Poisson
process of distribution points (as a function of the intensities of
the Poisson processes of subscribers and concentration points)
that minimizes this cost function. They have also extended the
above results for non-purely hierarchical models and have
derived the optimal intensity of Poisson process of distribution
points numerically, given the intensities of other two processes.
They have then generalized the cost function for networks with
more than three levels.
The algorithm proposed in this paper is similar to the
clustering algorithm in [6]. In [6], the authors have assumed
that the sensors are equipped with the capability of tuning the
power at which they transmit and they communicate with
power enough to achieve acceptable signal-to-noise ratio at the
receiver. We, on the other hand, assume a network in which the
sensors are very simple and all the sensors transmit at a fixed
power level; data between two communicating sensors not
within each other’s radio range is forwarded by other sensors in
the network. The authors, in [6], have observed in their
simulation experiments that in a network with one level of
clustering, there is an optimal number of clusterheads that
minimizes the energy used in the network. In this paper, we
have used the results provided in [11] to obtain the optimal
number of clusterheads at each level of clustering analytically,
for a network clustered using our algorithm to generate one or
more levels of clustering.
III. A NEW, ENERGY-EFFICIENT, SINGLE-LEVEL
CLUSTERING ALGORITHM
A. Algorithm
Each sensor in the network becomes a clusterhead (CH)
with probability
p
and advertises itself as a clusterhead to the
sensors within its radio range. We call these clusterheads the
volunteer clusterheads. This advertisement is forwarded to all
the sensors that are no more than
k hops away from the
clusterhead. Any sensor that receives such advertisements and
is not itself a clusterhead joins the cluster of the closest
clusterhead. Any sensor that is neither a clusterhead nor has
joined any cluster itself becomes a clusterhead; we call these
clusterheads the
forced clusterheads. Because we have limited
the advertisement forwarding to
k
hops, if a sensor does not
receive a CH advertisement within time duration
t
(where
t
units is the time required for data to reach the clusterhead from
any sensor
k
hops away) it can infer that it is not within k
hops of any volunteer clusterhead and hence become a forced
clusterhead. Moreover, since all the sensors within a cluster are
at most
k
hops away from the cluster-head, the clusterhead can
transmit the aggregated information to the processing center
after every
t
units of time. This limit on the number of hops
thus allows the cluster-heads to schedule their transmissions.
Note that this is a distributed algorithm and does not demand
clock synchronization between the sensors.
The energy used in the network for the information
gathered by the sensors to reach the processing center will
depend on the parameters
p
and
k
of our algorithm. Since the
objective of our work is to organize the sensors in clusters to
minimize this energy consumption, we need to find the values
of the parameters
p
and
k
of our algorithm that would ensure
minimization of energy consumption. We derive expressions
for optimal values of
p
and
k
in the next subsection.
B. Optimal parameters for the algorithm
To determine the optimal parameters for the algorithm
described above, we make the following assumptions:
a) The sensors in the wireless sensor network are
distributed as per a homogeneous spatial Poisson
process of intensity
λ
in 2-dimensional space.
b) All sensors transmit at the same power level and hence
have the same radio range
r
.
c) Data exchanged between two communicating sensors
not within each others’ radio range is forwarded by
other sensors.
d) A distance of
d
between any sensor and its
clusterhead is equivalent to
rd / hops.
e) Each sensor uses 1 unit of energy to transmit or receive
1 unit of data.
f) A routing infrastructure is in place; hence, when a
sensor communicates data to another sensor, only the
sensors on the routing path forward the data.
g) The communication environment is contention- and
error-free; hence, sensors do not have to retransmit any
data.
The basic idea of the derivation of the optimal parameter
values is to define a function for the energy used in the network
to communicate information to the information-processing
center and then find the values of parameters that would
minimize it.
0-7803-7753-2/03/$17.00 (C) 2003 IEEE IEEE INFOCOM 2003
1) Computation of the optimal probability of becoming a
clusterhead:
As per our assumptions, the sensors are distributed
according a homogeneous spatial Poisson process and hence,
the number of sensors in a square area of side
a2 is a Poisson
random variable,
N with mean A
λ
, where
2
4aA = . Let us
assume that for a particular realization of the process there are
n sensors in this area. Also assume that the processing center
is at the center of the square. The probability of becoming a
clusterhead is
p
; hence, on average, np sensors will become
clusterheads. Let
i
D be a random variable that denotes the
length of the segment from a sensor located at
ni
i
y
i
x ,...,2,1),,( = to the processing center. Without loss of
generality, we assume that the processing center is located at
the center of the square area. Then,
adA
a
yxnNDE
A
iii
765.0
4
1
]|[
2
22
==
∫
+=
. (1)
Since there are on an average np CHs and the location of
any CH is independent of the locations of other CHs, the total
length of the segments from all these CHs to the processing
center is
npa765.0
.
Now, since a sensor becomes a clusterhead with
probability
p
, the clusterheads and the non-clusterheads are
distributed as per independent homogeneous spatial Poisson
processes PP1 and PP0 of intensity
λλ
p=
1
and
λλ
)1(
0
p−= respectively.
For now, let us assume that we are not limiting the
maximum number of hops in the clusters. Each non-cluster-
head joins the cluster of the closest clusterhead to form a
Voronoi tessellation [10]. The plane is thus divided into zones
called the Voronoi cells, each cell corresponding to a PP1
process point, called its nucleus. If
v
N is the random variable
denoting the number of PP0 process points in each Voronoi
cell and
v
L is the total length of all segments connecting the
PP0 process points to the nucleus in a Voronoi cell, then
according to results in [11],
1
0
][]|[
λ
λ
=≈=
vv
NEnNNE (2)
2/3
1
0
2
][]|[
λ
λ
=≈=
vv
LEnNLE . (3)
Define
1
C to be the total energy used by the sensors in a
Voronoi cell to communicate one unit of data to the
clusterhead. Then,
r
nNLE
nNCE
v
]|[
]|[
1
=
==. (4)
Define
2
C to be the total energy spent by all the sensors
communicating 1 unit of data to their respective clusterheads.
Because, there are np cells, the expected value of
2
C
conditioned on N , is given by
]|[]|[
12
nNCnpEnNCE ===. (5)
If the total energy spent by the clusterheads to communicate
the aggregated information to the processing center is denoted
by
3
C , then,
r
npa
nNCE
765.0
]|[
3
==. (6)
Define C to be the total energy spent in the system. Then,
.
765.0
2
)1(
]|[]|[]|[
2/3
32
r
npa
p
p
r
np
nNCEnNCEnNCE
+
−
=
=+===
λ
(7)
Removing the conditioning on N yields:
.
765.0
2
1
765.0
2
1
][
]]|[[][
+
−
=
+
−
=
==
r
pa
pr
p
A
r
pa
pr
p
NE
nNCEECE
λ
λ
λ
(8)
][CE
is minimized by a value of
p
that is a solution of
01
2/3
=−− pcp . (9)
The above equation has three roots, two of which are
imaginary. The second derivative of the above function is
positive for the only real root of (9) and hence it minimizes the
energy spent.
The only real root of (9) is given by
0-7803-7753-2/03/$17.00 (C) 2003 IEEE IEEE INFOCOM 2003
2
3
3
1
22
3
1
22
3
2
1
.
3
)42733272(
)42733272(3
2
3
1
+++
+
+++
+
=
c
ccc
cccc
c
p
(10)
where
λ
ac 06.3= .
2) Computation of the maximum number of hops allowed
from a sensor to its clusterhead:
Till now we have not put any limit on the number of hops
(
k ) allowed between a sensor and its clusterhead. Our main
reason for limiting
k was to be able to fix a periodicity for the
clusterheads at which they should communicate to the
processing center. So, if we can find the maximum possible
distance (call it
max
R ) at which a PP0 process point can be
from its nucleus in a Voronoi cell, we can find the value of
k
by assuming that a distance
max
R from the nucleus is
equivalent to rR /
max
hops. Setting rRk /
max
= will also
ensure that there will be very few forced clusterheads in the
network.
Since it is not possible to get a value of
max
R such that we
can say with certainty that any point of PP0 process will be at
the most
max
R distance away from its nucleus in the Voronoi
Tessellation, we take a probabilistic approach; we set
max
R to a
value such that the probability of any point of PP0 process
being more than
max
R distance away from all points of PP1
process is very small. Using this value of
max
R , we can get the
value of parameter
k that would make the probability of any
sensor being more than
k
hops away from all volunteer
clusterheads very small.
Let
M
ρ
be the radius of the minimal ball centered at the
nucleus of a Voronoi cell, which contains the Voronoi cell. We
define
R
p to be the probability that
M
ρ
is greater than a certain
value
R
, i.e. )( RPp
MR
>=
ρ
. Then, it can be proved
that )09.1exp(7
2
11
Rp
R
λ
−≤ [11]. If
α
R is the value of R
such that
R
p is less than
α
, then,
λ
α
α
1
)7/ln(917.0
p
R
−
≤ . (11)
This means that the expected number of sensors that will
not join any cluster is
α
n if we set
=
−
λ
α
1
1
)7/ln(917.01
pr
k . (12)
To ensure minimum energy consumption, we will use a
very small value for
α
, which implies that the probability of
all sensors being within k hops from at least one volunteer
clusterhead is very high.
For 001.0=
α
and values of
p
and k computed according
to (10) and (12), for a network of 1000 sensors, on an average 1
sensor will not join any volunteer clusterheads and will become
a forced clusterhead. The optimal value of
p
for a network
with 1000 nodes in an area of 100 sq. units is 0.08, which
means 80 nodes will become volunteer clusterheads on an
average. Hence, for a network of 1000 nodes in an area of 100
sq. units, only 1.23 % of all clusterheads are forced
clusterheads.
C. Simulation Experiments and Results
We simulated the algorithm described in Section III for
networks with varying sensor density ( d ) and different values
of the parameters
p
and
k
. In all these experiments, the
communication range of each sensor was assumed to be 1 unit.
Fig. 1 shows the output of one of these simulations of our
algorithm with parameters
p
and
k
set to 0.1 and 2 on a
network of 500 sensors distributed uniformly in a square area
of 100 square units.
To verify that the optimal values of the parameters
p
and
k
of our algorithms computed according to (10) and (12) do
minimize the energy spent in the system, we simulated our
clustering algorithm on sensor networks with 500, 1000 and
2000 sensors distributed uniformly in a square area of 100 sq.
units. Without loss of generality, it is assumed that the cost of
transmitting 1 unit of data is 1 unit of energy. The processing
center is assumed to be located at the center of the square area.
For the first set of simulation experiments, we considered a
range of values for the probability (
p
) of becoming a
clusterhead in the algorithm proposed in Section III. For each
of these probability values, we computed the maximum number
of hops (
k
) allowed in a cluster using (12) and used these
values for the maximum number of hops allowed in a cluster in
the simulations. The results of these simulations are provided in
Fig. 2. Each data point in Fig. 2 corresponds to the average
energy consumption over 1000 experiments. It is evident from
Fig. 2 that the energy spent in the network is indeed minimum
at the theoretically optimal values of the parameter
p
computed using (10) (let us call this optimal value
opt
p ),
which are given in Table I for 500, 1000 and 2000 sensors in
the network.
0-7803-7753-2/03/$17.00 (C) 2003 IEEE IEEE INFOCOM 2003
