Fuzzy C-Means Clustering Protocol for Wireless

Sensor Networks
D.C. Hoang, Student Member, IEEE, R. Kumar, Member, IEEE, and S.K. Panda*, Senior Member, IEEE
*Department of Electrical and Computer Engineering
National University of Singapore
Singapore 117576
Telephone: (65) 6516-6484
Email: eleskp@nus.edu.sg

Abstract—Hierarchical or cluster based routing methods sensing task and only send the information to the CH when
are well-known techniques with special advantages related to neccessary, while CHs collect data from other nodes and send
scalability and efficient communication. Clustering technique is to the end users.
also utilized as an energy-efficient routing in Wireless Sensor
Networks where the nodes with the highest residual energy can - In location-based protocol, routing data needs the infor-
be used to gather data and send the information. In this paper, mation of sensor nodes locations in the deployed field.
we proposed and analyzed a cluster based protocol using Fuzzy Among these, hierarchical or cluster-based protocols are
C-Means (FCM) method. A comparison among FCM, K-Means well-known techniques with special advantages related to
clustering, LEACH and MTE has been done. Simulation result scalability and efficient communication. The concept of hi-
shows that FCM protocol can reduce energy consumption and
improve the network lifetime. erarchical routing is also utilized to perform energy-efficient
Index Terms—Hierarchical routing, clustering, Fuzzy C- routing in WSNs. By grouping the nodes into clusters as
Means, Wireless Sensor Networks. shown in Fig.1 with the assistance of data aggregation and
fusion techniques, efficient usage of energy resource is ob-
I. I NTRODUCTION tained because the overall amount of data transmitted to the
Wireless Sensor Networks (WSNs) consist of hundreds to BS is significantly decreased, intra-cluster communication
thousands of tiny sensor nodes equipped with sensing, data enables to reduce the transmission distance of non-CH nodes
processing, and communication units [1]. These sensor nodes and then reduce energy consumption. Furthermore, duty
are used to collect information about ambient environment, cycling of the non-CH nodes can be carried out by the CH
eg. temperature, humidity, light, vibration, acoustic, etc. The within the cluster, therefore, member nodes are allowed to
measurement is pre-processed and useful data is transferred enter sleep mode for a longer time. However, gathering and
to other sensor nodes or to a Base Station (BS). Due to processing information as well as transferring data to the BS
these capabilities, WSNs can be applied in various poten- cause higher energy dissipation at the CH, and thus lead to
tial applications such as target tracking, habitat monitoring, shorter lifetime of the CH and the network connection cannot
healthcare monitoring, surveillance, etc. However, to make be maintained anymore. In order to balance the energy usage
WSNs feasible to be employed, a number of requirements in of sensor nodes and CH, the solution of rotating CH roles
the design and operation of the network need to be satisfied. among sensor nodes within the cluster is popularly used.
Since sensor nodes are powered by limited energy source, The selection of CH is based on residual energy of the sensor
energy conservation is commonly considered the most key nodes, any nodes with residual energy higher thha a threshold
challenge in order to guarantee the connectivity of the net- can be chosen to become CH.
work and extend the lifetime of the sensor nodes, especially Low Energy Adaptive Clustering Hierarchy (LEACH) [3]
when the deployment field is inaccessible and battery cannot is a typical cluster-based protocol using a distributed cluster-
be replaced. Even if unlimited energy source like solar, wind, ing formation algorithm. The cluster heads are selected with
etc., is utilized, efficient operation of sensor networks is a predetermined probability, other nodes choose the nearest
necessary because of the fluctuation and intermittent nature cluster to join, based on the strength of the advertisement
of these sources. It is also recognized that usually commu- message they received from the cluster heads. After forming
nication task consumes the most energy during the network the clusters, cluster heads compress data arriving from the
operation. Many routing protocols have been proposed to sensor nodes and send an aggregated packet to the BS in
obtain efficient-energy communication for the WSNs for order to reduce the amount of information sent to the BS.
recent years. In [2], routing techniques are classified into To avoid intracluster and intercluster collision, Time Division
three categories based on network structure which are flat, Multiple Access (TDMA) or Code-Division Multiple Access
hierarchical and location-based routing protocols: (CDMA) can be utilized, thus a better efficient communi-
- In flat routing protocol, all sensor nodes are typically cation is achieved. LEACH has an outperformed operation
assigned equal roles and has the same functionality; sensor compared with Direct Communication where the sensor
nodes collaborate to perform the sensing task as well as nodes send data directly to the base station and Minimum
communication task. Transmission Energy (MTE) routing protocol [4]. In the case
978-1-4244-6392-3/10/$26.00 ©2010 IEEE 3477
- In hierarchical protocol or cluster based routing method, of Direct Communication, the sensor nodes send data to the
there are usually two types of sensor node: Cluster Head BS directly, it is shown that the further nodes from the BS
(CH) and non-CH nodes. Non-CH nodes mainly carry out consume more energy to transmit data and thus have a shorter
life than that of nodes closer to the BS. Meanwhile, MTE
protocol routes messages from a node through its nearest
neighbor toward the BS, in this case, nodes nearer to the BS
have to relay a huge amount of data, thus they are run out of
energy before the further nodes [3]. The simulation result
shows that LEACH can considerably extend the network
lifetime in terms of duration of time until first node dies,
and guarantee the coverage of the network for a longer time
until the last node dies.
degree of belonging to each cluster rather than hard partition-
ing them into only one cluster. Fuzzy C-Means algorithms
assists to optimize the clusters based on minimizing the
distance between the sensor node and the cluster center. This
method guarantees a uniform creation of clusters in randomly
deployed sensor networks, clusters are the groups with
high density of sensor nodes, thus the total spatial distance
between the cluster head and sensor nodes is significantly
reduced. Traffic load is balanced among the cluster heads
in the global network and energy consumption is balanced
among sensor nodes in the local cluster.
The structure of the paper is organized as follow. The
network operation assumptions and radio energy model are
described in section II. Section III presents fuzzy c-means
clustering protocol. The simulation result and analysis of the
protocol are shown in Section IV and discussed in Section
V respectively, and the conclusion is given in Section VI.
II. P RELIMINARIES
We first present the assumptions and model of the network
under consideration.

Fig. 1. Architecture of the Hierarchical Network A. Assumptions

Although, the lifetime of the sensor nodes and the network We consider scenario of application in which sensor nodes
is significantly increased, there are still some issues with are deployed randomly in order to continuously monitor the
LEACH. The random selection of the CH may obtain a environment. The information collected by sensor nodes is
poor clustering set-up, and CHs may be redundant for some sent to a base station located far from the deployment field.
rounds of operation. Furthermore, the distribution of CHs Furthermore, other assumptions are made as follow:
is not uniform, thus some sensor nodes have to transfer - Sensor nodes as well as base station are stationary after
data through a longer distance. In [5], LEACH-centralized being deployed in the field.
(LEACH-C) is proposed as an improvement of LEACH - The network is considered homogeneous and all of the
which uses a centralized clustering algorithm to form the sensor nodes have the same initial energy.
clusters. The purpose of LEACH-C is to produce a better - Each sensor node knows their own geographical position.
clusters by dispersing the cluster head nodes throughout - All nodes measure the environmental parameters at a
the network. During the set-up phase, the BS collects the fixed rate and send it periodically to the receiver nodes.
information of the position and energy level from all sensor - The radio channel is symmetric such that energy con-
nodes in the networks. Using this information, the BS finds sumption of data transmission from node A to node B is the
the number of cluster heads and configures the network same as that of transmission from node B to node A.
into clusters. The amount of energy of transmission data Each sensor nodes can operate either in sensing mode to
between the non-cluster heads to CH is saved by the means monitor the environment parameters and transmit to the base
of minimizing the total squared distances between all the station or cluster head mode to gather data, compress it and
non-cluster head nodes and the closest cluster head. forward to the BS.
Other cluster-based protocols based on LEACH are PE- B. Radio energy models
GASIS [6], TEEN [7] and APTEEN [8]. However these
As mentioning above, in most cases of application, among
protocols improve the network lifetime by enhancing the
its components consuming power such as data processing
efficiency of data transmission but the cluster structure is
unit, sensing unit, memory storage and transceiver, the ra-
not optimized.
dio communication unit takes the most energy expenditure.
Another technique popularly used to cluster sensor net-
Therefore, we mainly consider the energy consumption of the
work is k-means. In [9], the authors proposed a Balanced
sensor nodes for communication within the network. Sensor
Parallel K-means (BPK-means) based clustering protocol
node main components and associated energy consumption
based on k-means algorithm. BPK-means protocol is de-
are shown in Fig.2.
veloped to reduce energy consumption of the sensor node
during communication with the CHs. However, with the
random deployment of the sensor nodes in an unattended
field, hard partitioning the network by means of using k-
means algorithm may result in vagueness of classifying
sensor nodes near the boundary of the clusters, thus, the
optimization of cluster formation is not obtained.
3478
In this paper, we developed and analyzed an innovative
cluster-based protocols using Fuzzy C-Means algorithms
where the sensor nodes are grouped into clusters with a Fig. 2. Sensor Node Components and Energy Consumption
for Data Aggregation and Communication Task
 Both the free space and multipath fading channel models III. FCM CLUSTERING PROTOCOLS
in [10],[5] are used to compute energy dissipated during FCM clustering protocols is centralized clustering algo-
the process of transmitting and receiving information. The rithms, the base station computes and allocates sensor nodes
energy consumption for transmitting a l bit message over a into clusters according to the information of their location
distance d is and the cluster head is assigned to the node having the
largest residual energy. We consider a network of N sensor
ET x = Eelec ∗ l + Ef s ∗ l ∗ d2 , d < d0 (1)
4
nodes which is partitioned into c clusters: C1 , C2 , ..., Cc .
ET x = Eelec ∗ l + Emp ∗ l ∗ d , d ≥ d0 (2) The purpose of the cluster formation in this protocol is to
minimize the following objective function:
and for receiving this message respectively is:
c X
N
ERx = Eelec ∗ l (3)
X
Jm = um 2
ij dij (9)
i=1 j=1
where Eelec is the energy spent to operate the transceiver
circuit, Ef s and Emp are the energy expenditure of trans- where
mitting one bit data to achieve an acceptable bit error rate uij is node j’s degree of belonging to cluster i
and is dependent on the distance of transmission in the dij is the distance between node j and the center point of
case of free space model and multipath fading model. If cluster i
The degree uij of node j respected to cluster is calculated
the transmission distance is less than a threshold d0 , the
and fuzzyfied with the real parameter m > 1 as below:
free space model is applied; otherwise, we use the multipath
model. The threshold d0 is calculated as 1
uij = Pc 2
dij m−1
(10)
k=1 ( dkj )
q
d0 = Ef s /Emp (4)
The distance between the sensor node and the center point
Another parameter is also taken into account is the data is Euclidean distance. By achieving minimization of the
aggregation energy expenditure which is set as Eda = spatial distance, the energy balance among sensor nodes is
5nJ/bit/message. We use the following values for other optimized.
parameters: Eelec = 50nJ/bit, Ef s = 10pJ/bit/m2 , and Our FCM clustering protocol include 3 phases: clustering
Emp = 0.0013pJ/bit/m4 [5]. calculation, cluster head selection and data transmission. The
Considering a N node network partitioned into c clusters, operation of the protocol is partitioned into rounds. In each
the average number of nodes in a cluster is N/c, the energy round, the cluster heads collect data from all cluster members
consumption of the CH to receive message from the non- and transfer to the BS.
cluster head is A. Clustering calculation
N We consider the application scenario that N sensor nodes
ERx−CH = lEelec ( − 1) (5)
c are deployed randomly into a field with an area of M ×M m2 .
and to aggregate data into a l-bit message and send it to the After being spread out, these sensor nodes send a HELLO
BS is message to the base station with the information of their
N geographical location; based on this information the base
ET x−CH = lEda + lEelec + lEmp d4toBS (6) station will calculate the cluster centers and allocated sensor
c
nodes into cluster using FCM algorithm. FCM algorithm is
where dtoBS is the average distance from one CH to the BS. first proposed by Bezdek [11] to be used in cluster analysis,
Assuming that the distance from the non cluster head node pattern recognition, image processing. In the case of our
and the CH is short, the energy consumed by the non-cluster application, FCM algorithm is applied to cluster the sensor
head node to transmit a l-bit message is nodes. Each node is assigned a degree of belonging to cluster
head rather than completely being a member of just one
ET x−nonCH = lEelec + lEf s d2toCH (7) cluster. Therefore, the nodes close to the boundary of a
cluster may become members of the cluster with a degree ap-
1 M2
where dtoCH ≈ 2π c [5] which is the average distance proximating the degree of belonging to the neighbor clusters.
from one node to its CH and M is the network diameter. The calculation is carried out by following the Algorithm.1
Thus, the total energy dissipated in the network during a The convergence is achieved when the difference between
round of collecting data and transmitting to BS is the coefficients in two iterations is less than a threshold or
a large number of iterations is reached.
Eround = l (2N Eelec + N Eda + cEmp d4toBS After forming the clusters, the BS chooses the nearest
+N Ef s d2toCH ) nodes to cluster centers to become CH. Once the cluster
= l (2N Eelec + N EDA + cEmp d2toBS creation is complete, base station send the information of
1 M2 the cluster head and to which cluster a node belongs to all
+N Ef s ) (8) of the nodes.
2π c
To identify the number of clusters, the following formula
This total energy expenditure consists of the average [5] can be used:
3479
energy dissipated by data transmission of non-cluster head √ s
nodes and CHs and the energy consumption for data collec- N Ef s M
copt = √ 2 (12)
tion and fusion of the CHs. 2∗π E mp dtoBS
Algorithm 1 FCM algorithm for cluster formation IV. S IMULATION RESULT AND D ISCUSSION
for j = 1 to N do A. Experiment 1 - Network lifetime assessment with different
Node j is given the coefficient uij for being a member protocols
of cluster i
In the first experiment, FCM clustering protocol is eval-
end for
uated by simulating a 300 node network with MATLAB.
repeat
Sensor nodes are deployed in an area of 250×250m2 shown
for i = 1 to c do
in Fig. 3, the BS is located at (125m,-75m). The number of
Compute the centroid of each cluster
Pn clustering is 5 which is calculated by using the Equation. 12
m
j=1 uj pos(node j) with the average dtoBS = 100m. Each sensor node transmits
pos(centeri ) = Pc (11)
j=1 uj
m a 4000 bit message with 150 bit header to the cluster head at
each round. The initial energy supplies to each sensor nodes
end for is 2J.
until the algorithm is converged

which is found from the total energy dissipated in one round

given in Equation.8
B. Cluster head selection
After the cluster is created, the non-cluster head nodes
send data toward the base station through the cluster heads.
The process of selecting clusters is repeated every round
of exchanging data among sensor nodes. Only at the first
stage, the cluster head of each cluster is chosen by the base
station; after that the current cluster head makes decision of
selecting which node will become the cluster head at the
next round. During the transmission from the sensor nodes
to CH, residual energy of each nodes are attached to the data Fig. 4. Cluster formation with LEACH protocol at a
packet, this information assists the CH choose the node with arbitrary round
the highest residual energy and nearest to the cluster center
to be cluster head at the next round. Based on the number
of the alive nodes within the cluster, the new CH creates a
TDMA schedule to allocate the time when cluster members
can transmit.

Fig. 5. Cluster formation with FCM protocol

Fig.4 and Fig. 5 show the cluster structure of the network

when using LEACH and FCM clustering protocols respec-
tively. It is found that the distribution of clusters in LEACH
is not uniform, some clusters consist of a huge number of
nodes spread out in a large area wheareas the others have a
Fig. 3. Deployment of sensor nodes into monitored field
few as shown in Fig.4.
C. Data transmission Hence, in some rounds, the CH of the group with huge
Once the CHs are selected and the transmission scheduled number of nodes has to suffer from heavy traffic load which
is made, the sensor nodes start to transmit data to the CHs. causes significant power consumption of the CH and high
Transmission power of non-cluster head nodes is optimized chance of wasting energy due to data collision.
because of the minimum spatial distance to the CHs achieved Moreover, many sensor nodes do not have the CH in their
by FCM algorithm. Furthermore, as TDMA scheduling pro- proximity, so more energy is required to transmit data from
tocol is used, non-cluster head nodes only need to turn on these nodes to CHs over longer distance. Meanwhile by
their radio component during the transmission, and can turn using FCM the clusters has better formation where mean
3480
off after finish transmitting. Data aggregation and fusion is distance from each node to the cluster is minimized. It is
done at the CHs, thus the amount of information is reduced, more efficient to balance the load of network and to distribute
the CHs only send the compressed data to the BS. the nodes among clusters by using FCM.
 Fig. 6. Distribution of dead nodes (dots) with Direct Fig. 7. Distribution of dead nodes (dots) with MTE
Communication after 300 rounds after 300 rounds

Fig. 8. Distribution of dead nodes (dots) with Fig. 9. Distribution of dead nodes (dots) with FCM
LEACH after 700 rounds after 1600 rounds

TABLE I
D URATION OF TIME UP TO THE FIRST NODE DIES IN THE NETWORK

Protocols Direct MTE LEACH K-Means FCM

Lifetime of the 11 33 541 1396 1598
first dead node
(rounds)

In Fig.10, the number of alive nodes over the operating

time of the network by using different protocols is compared.
In our work, the performance of Direct Communication,
MTE, LEACH, K-means and FCM are studied. The different
duration of time up to the first dead node by applying
different protocols is given in Table.I. It is obviously seen
that the lifetime of the network with 100% nodes alive when
Fig. 10. Number of node alive over the time with
different cluster based protocols
using FCM protocol is much longer than the network lifetime
when LEACH is employed, and also the lifetime of network
with K-Means protocol.
The distribution of dead nodes is presented in Fig.6, Fig.7,
Fig.8 and Fig.9. As mentioned in Section I and proven in B. Experiment 2 - Energy expenditure evaluation within the
[4], in Direct Communication, nodes which are further from deployment network
the BS deplete energy faster as the transmitting distance is We study the average energy dissipated within the network
further; in MTE, nodes which are closer to the BS run out in the second experiment. 100 sensor nodes are randomly
of energy earlier because of forwarding a huge amount of deployed in areas with different diameters with 1J initial
data received from the further nodes. energy of each node. The simulation is run to observe
In LEACH, although the nodes further from the BS have the changing of energy consumption with the variation of
higher chance to run out of energy first, the energy balance network diameter and different values of electronic energy.
among the network assist LEACH network obtains a larger Fig.11 shows the average energy expenditure of network
covered area after a longer time of operation, 700 rounds with different protocols over the diameter after 200 rounds.
compared with 300 rounds of operation in the case of using With the small network diameter, energy expenditure of
Direct Communication and MTE. Meanwhile, when FCM is the network by using different protocols are almost the
3481
employed, the coverage of the network upto 1600 rounds is same. However, when the network diameter increases, FCM
much higher, thus, it guarantees the larger area of the field outperforms MTE, Direct Communication, LEACH and K-
sensed by sensor nodes. Means. This is because of the better cluster distribution and
traffic load balance among the network achieved by using
FCM as mentioned above.
Fig. 11. Average energy dissipated within the
network over the network diameter after 200 rounds
Fig. 12. Average energy dissipated within the
network by using MTE and FCM over the network
diameter and electronics energy after 200 rounds
Fig. 13. Average energy dissipated within the
network by using LEACH and FCM over the network
diameter and electronics energy after 200 rounds
3482
Fig. 14. Average energy dissipated within the net-
work by using K-Means and FCM over the network
diameter and electronics energy after 200 rounds
Another energy expenditure comparison made with the
same size of network among different protocols is shown
in Fig.12, Fig.13 and Fig.14. Once the electronics energy or
the network diameter increase, FCM network consumes less
energy than the network with MTE in Fig.12, with LEACH
in Fig.13 and with K-Means in Fig.14. The improvement of
FCM is much clearer when the deployment field is larger
and the electronic energy grows.
V. C ONCLUSION
In this paper, we present a centralized cluster-based pro-
tocol, FCM clustering. This protocol uses FCM algorithm
to create cluster structure in order to minimize the spatial
distance among the sensor nodes and thus a better cluster
formation is obtained. With the facilitation of data aggrega-
tion, cluster head rotation and TDMA schedule techniques
in clusters, energy consumption is balanced among all the
sensor nodes and the amount of data transmitted to the BS
is reduced remarkably. Our simulation results show that by
applying FCM clustering algorithm the power consumption
is reduced and the life time of the network is extended
significantly when compared with LEACH and K-Means.
VI. ACKNOWLEDGMENTS
This research is supported in part by a grant from A-Star
in IEDS project.
R EFERENCES
[1] I. F. Akyildiz and W. Su and Y. Sankarasubramaniam and E. Cayirci,
Wireless sensor networks: a survey, Computer Networks, vol 38, pp
393-422, 2002.
[2] Jamal N. Al-Karaki and Ahmed E. Kamal, Routing Techniques in
Wireless Sensor Networks: A Survey, IEEE Wireless Communications,
pp.6-28, 2004.
[3] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, Energy-
efficient communication protocols for wireless sensor networks, In
Proceedings of the 33rd Annual Hawaii International Conference on
System Sciences (HICSS), Hawaii, USA, January 2000.
[4] M. Ettus, System capacity, latency, and power consumption in multihop-
routed SS-CDMA wireless networks, Radio and Wireless Conference
(RAWCON), pp. 55-58, August 1998.
[5] W. B. Heinzelman, A. P. Chandrakasan, and H. Balakrishnan, An
application-specific protocol architecture for wireless microsensor net-
works, IEEE Transactions on Wireless Communications, vol. 1, pp.
660670, 2002.
[6] S. Lindsey and C. Raghavendra, PEGASIS: Power-Efficient Gathering
in Sensor Information Systems, IEEE Aerospace Conf. Proc., vol. 3,
916, pp. 112530, 2002.
[7] A. Manjeshwar and D. P. Agarwal, TEEN: a Routing Protocol for
Enhanced Efficiency in Wireless Sensor Networks, Parallel and
Distributed Processing Symposium., Proceedings 15th International, pp.
2009-2015, 2001.
[8] A. Manjeshwar and D. P. Agarwal, APTEEN: a hybrid protocol for
efficient routing and comprehensive information retrieval in wireless
sensor networks, Parallel and Distributed Processing Symposium.,
Proceedings International, pp. 195-202, 2002.
[9] Liansheng Tan, Yanlin Gong and Gong Chen, A balanced parallel clus-
tering protocol for wireless sensor networks using K-means techniques,
hskip 1em plus 0.5em minus 0.4emProceedings of The Second Inter-
national Conference on Sensor Technologies and Applications (SEN-
SORCOMM 2008), Cap Esterel, France, pp. 25-31, August 2008.
[10] T. Rappaport, Wireless Communication: Priciples and Practice, En-
glewood, NJ: Prentice Hall, 1996.
[11] Bezdek, J.C, Pattern Recognition with Fuzzy Objective Function
Algorithms, Plenum Press, New York, 1981.