INTRODUCTION
A sensor network is an infrastructure composed of sensing (measuring), computing
and communication elements that enable administrators to observe and react to
events and phenomena in a specified environment (Sohraby et
al., 2007). Wireless Sensor Networks (WSNs) are used in airtraffic
control, battlefield management, commercial application, food safety, intrusion
detection and many other fields (Puccinelli and Haenggi,
2005; Akyildiz and Vuran, 2010). Each sensor node
in a WSN is composed of a sensor, an A/D converter, a processor with memory
(e.g., a StrongARM SA1100 processor), RF circuits, a DC converter and a battery,
as shown in Fig. 1. Each sensor node has a very limited energy
supply and may not be rechargeable. In some cases, nodes may be selfpowered.
Energy is used for sensing, data processing and communication. Communication
is the most energyintensive process, consuming more energy than the computation
process in the operating system (Akyildiz and Vuran, 2010).
Therefore, a sensor node must be designed intelligently taking all its aspects
into consideration, including architecture, algorithms, protocols and circuitry.
The current study is focused on the design of an algorithm for sensor nodes
that can prolong the lifetime of a WSN.
A WSN is a largescale network that must be arranged in an appropriate manner,
especially for data aggregation. The data aggregation process is intended to
balance the load and thereby to extend the network lifetime. The nodeclustering
algorithm has been proven to be an effective way of organizing a network into
a connected hierarchy. In a nodeclustering algorithm, several nodes are aggregated
to form a group. This scheme usually operates in two phases: node cluster setup
and cluster maintenance (Zhang et al., 2009).
In the nodecluster setup phase, Cluster Heads (CHs) are selected among the
nodes in the network using various selection schemes, as shown in Table
1. After the CHs have been selected, other nodes that are affiliated with
each CH form clusters. Nodes that are not a CH are called nonCH nodes. Each
CH, acting as a router, transmits data collected from nonCH nodes to the sink
node. In the cluster maintenance phase, the cluster configuration may be changed
after the initial cluster is set up because of node movements or topology changes.
An example of a nodecluster structure is shown in Fig. 2.
The cluster head node will lose more energy than a noncluster head node because
it transfers data over longer distances. Hence, to distribute the load uniformly
among all the nodes, the network must recluster itself periodically, selecting
energyabundant nodes to serve as CHs. Thus, the network will achieve energy
efficiency, reduce channel contention and reduce packet collisions, resulting
in better network throughput under high load (Younis and
Fahmy, 2004).
Node clustering has been studied comprehensively in the data processing and wired network literatures.
Table 1: 
Different clusterhead (CH) selection schemes. 


Fig. 2: 
Example of a nodeclustering structure in a WSN 
Unfortunately, the clustering approaches developed and implemented in these
areas cannot be used directly in WSNs because of the unique deployment and operational
characteristics of these networks. The first study of clusterbased networks
was carried out by (Heinzelman et al., 2002).
They proposed a Lowenergy Adaptive Clustering Hierarchy (LEACH), a routing
protocol with selforganizing and adaptive clustering behavior that randomly
distributes the energy load among the sensors in a network. LEACH uses localized
coordination to enable scalability and robustness for active networks. Moreover,
it also uses data fusion in the network to reduce the amount of information
that must be sent to the sink node. This study was followed by several attempts
to implement a nodeclustering algorithm in a WSN, including Hybird Energyefficient
and Distributed (HEED) (Younis et al., 2006),
Powerefficient Gathering in Sensor Information System (PEGASIS) (Lindsey
and Raghavendra, 2002), Treebased Energy Efficient Protocol for Sensor
Information (TREEPSI) (Satapathy and Sarma, 2006) and
Advanced LEAH (ALEACH) (Ali et al., 2008).
Most of the clustering algorithms named above assume that all sensor nodes
are equipped with the same quantity of links, computational resources and energy;
this is the case in homogeneous sensor networks. However, this kind of network
cannot be assumed because different sensor nodes do not always have the same
communication and sensing capabilities. The fact that sensor nodes use the same
platform does not guarantee that they will have exactly the same physical properties
(Das and Ammari, 2009). In contrast with homogeneous
networks, there are also heterogeneous networks, called here Heterogeneous Wireless
Sensor Networks (HWSNs). Recently, researchers have shown an increased interest
in heterogeneous network research because of its ability to extend the lifetime
of networks (Katiyar et al., 2011), improve the
reliability of data transmission (Yarvis et al.,
2005) and decrease the latency of data transfer (Yu
et al., 2007). A typical heterogeneous WSN consists of a large number
of normal nodes and a few heterogeneous nodes (DuarteMelo
and Liu, 2002; Yu et al., 2007). The normal
node, whose main tasks are to sense and to issue data reports, is inexpensive
and sourceconstrained. The heterogeneous nodes which provide data filtering,
fusion and transport, are more expensive and more capable. A considerable amount
of experimental research on HWSNs has been done by Du et
al. (2008), who provided examples of two real sensor network deployments
that use heterogeneous nodes for processing and transport tasks. Furthermore,
many clustering algorithms have been proposed for an HWSN environment, for example,
Stable Election Protocol (SEP) (Smaragdakis et al.,
2004), Distributed Energy Efficient Clustering (DEEC) (Qing
et al., 2006), Energy Efficient Hetrogenous Clusters (EEHC) (Kumar
et al., 2009) and others. This study presents a deeper explanation
of this entire algorithm and a comparison with other algorithms.
NODECLUSTERING ALGORITHM
Recently, researchers have devoted considerable attention to nodeclustering
algorithms because of their ability to prolong the lifetime of WSNs.

Fig. 3: 
Architecture of a wireless sensor network 
All the necessary terms involved in designing nodeclustering algorithms are
explained in this section.
Terminology
Homogeneous vs. heterogeneous: In a homogeneous network, all nodes
in the WSN have the same storage, computation, communication, sensing and energy
capabilities. The communication links between the sensor nodes are symmetric.
As a result, in such a network, a pair of neighboring sensors can communicate
directly; otherwise, the network is heterogeneous. The combination of both is
called a hybrid network. The heterogeneous network is a fair assumption for
practical implementations.
Optimum number of clusters: The number of clusters required in a network
is closely related to achieving optimal energy consumption. In some published
approaches, the set of CHs is predetermined and thus the number of clusters
is preset. Randomly picking CHs from the deployed sensors usually yields a variable
number of clusters. Typically, use of at least 10 clusters is recommended (Heinzelman
et al., 2002).
Stability: Stability is influenced by how the numbers of clusters in
the network are counted. If the cluster counts vary and a node’s membership
evolves over time, this is called an adaptive scheme. Otherwise, the clustering
scheme is called fixed because the number of clusters does not change and sensor
nodes do not switch from one cluster to another in the network (Abbasi
and Younis, 2007).
Cluster head (CH): The CH acts as a router that receives sensing data
from all members of the cluster and transmits data to the sink node using single
or multihop data transmission. Multihop data transmission is the appropriate
technique for sink nodes located far away from the sensing area. Otherwise,
singlehop data transmission is better. Sometimes the CH can be designated as
a mobile or immobile node (Abbasi and Younis, 2007).
Noncluster head node: Each node can become the CH in any round. A node that is not selected as the CH becomes a nonCH node which then selects its corresponding cluster based on its signal strength with the current CH.
Transmission range: This term refers to a disk of a given radius, including its boundary and center. Transmission range is usually the threshold that differentiates two kinds of communication schemes in the network, namely free space and multipath fading models.
Singlehop vs. multihop: In contrast to multihop transmission, data in singlehop mode are transmitted from one node to another without using a third node as a router.
Mobile vs. immobile nodes: A node can either move within the network (mobile node) or be stationary (immobile node).
Figure 3 illustrates all the terminological definitions given above.
HETEROGENEOUS WIRELESS SENSOR NETWORKS (HWSNs)
Types of heterogeneous resources: Yarvis et al.
(2005) list three types of resource heterogeneity in sensor nodes. The first,
computational heterogeneity, means that at least one node contains a more powerful
microprocessor and larger memory than a normal node. The second type is known
as link heterogeneity and means that at least one heterogeneous node has higher
bandwidth and a longerdistance network receiver than a normal node. The third
type, energy heterogeneity, means that at least one heterogeneous node is linepowered
or that its battery is replaceable.
For the most part, energy heterogeneity is the most vital type of resource heterogeneity. A sensor network generally suffers a negative impact when both computational and link heterogeneity have consumed large quantities of energy resources. Therefore, energy heterogeneity must be given top priority when creating a heterogeneous environment in a WSN.
The impact of heterogeneous resources: The main benefit of a heterogeneous
network is extended network lifetime. Average energy consumption can be decreased
by forwarding packets from normal nodes to the sink node (Yu
et al., 2007). In addition, if each hop in the network can significantly
reduce its endtoend delivery time, this can improve the reliability of data
transmission because there will be fewer hops between the normal sensor nodes
and the sink. Consequently, use of computational or link heterogeneity can decrease
the latency of data transfer in the network.
OBJECTIVES OF HETEROGENEOUS CLUSTERING
Nodes are formed into clusters to attain certain objectives, as explained below:
Enhanced network lifetime: The lifetime of the network can be extended by reenergizing the sensor nodes. This will create heterogeneity because some of these nodes are equipped with more energy than the normal nodes already in use. Consequently, all the normal nodes can transmit sensed data to the sink node through the nearest heterogeneous node. This will make the average energy consumption for forwarding data to the sink node much less than in the homogeneous sensor network.
Load balancing: LEACH guarantees, in the homogeneous case, that any
unstable region in the network will be of limited size. No such guarantee exists
for heterogeneous networks. As a solution, a weighted probability has been introduced
to increase the stable region in a network which is implemented as a heterogeneous
network (Smaragdakis et al., 2004).
Decreasing the latency of data transfer: HWSNs with computational heterogeneity may decrease the processing latency for immediately adjacent nodes. Moreover, if link heterogeneity is implemented in the HWSNs, this will reduce the waiting time in the transmission queue.
Improving the reliability of data transmission: In a heterogeneous network,
the number of hops to transmit data from sensor nodes to the sink node is less
than in a homogeneous network (Yu et al., 2007).
For this reason, HWSNs can achieve much higher endtoend delivery speeds.
CLUSTERING ATTRIBUTES
Clustering attributes are a vital part of categorizing and differentiating
the various clustering algorithms which have been proposed for HWSNs. There
are three types of attributes of clustering algorithms (Abbasi
and Younis, 2007).
Cluster properties: Important cluster attributes include the optimal number of clusters, the stability of the network and the data transmission technique used for intracluster or intercluster connectivity. Normally, direct communication is selected for nearby data communication and multihop data communication for longdistance communication.
Cluster head capabilities and selection: The CH is a particularly powerful node in a clustering algorithm. It can be either mobile or immobile. When a CH is mobile, its cluster’s sensor membership changes dynamically and clusters will need to be continually maintained. If the CH is immobile, a good choice of CH yields a stable cluster which facilitates intra and intercluster network management. Each proposed algorithm has its own types of node heterogeneity, as was explained in the previous section. Furthermore, sometimes the CH acts as a data aggregation or relay node. All these tasks vary depending on the network implementation. Finally, the CH selection scheme varies depending on many factors such as the initial, residual and average energy of all the nodes in the network.
Clustering process: The methodology, the objective of node clustering and the clusterhead selection method are important factors that must be considered during the clustering process. For HWSNs, various methodologies can be selected, including adaptive clustering, energy multilevel clustering and stabilityoriented clustering. Consequently, the objectives of clustering must be considered with respect to load balancing, network connectivity and similar tasks. Furthermore, a clusterhead selection method is needed to determine the CH using stochastic or deterministic methods or a combination of both.
ENERGYEFFICIENT NODECLUSTERING ALGORITHM FOR HWSNs
Energy consumption model: Energy consumption by the sensor nodes in
WSNs can be calculated using the energy model proposed by Heinzelman
et al. (2002). The energy consumption for sending k bits of data
over a distance d can be calculated using Eq. 1. The distance
threshold d_{0} will differentiate the type of data communication. The
first equation is the freespace model, in which the transmission power is attenuated
to d^{2} for d<d_{0}. The second equation is the multipath
fading model, in which the transmission power is attenuated to d^{4}
for d>d_{0}:
where, ε_{frissamp} and ε_{tworayamp} are the twochannel model parameters for the energy needed for power amplification. Consequently, the energy consumed when receiving k bits of data can be calculated using Eq. 2:
where E_{elec} represents the energy consumed in transmitting or receiving a bit of data.
NODECLUSTERING ALGORITHM FOR HWSNs
Nodeclustering algorithms are a wellknown technique for reducing energy consumption in a sensor network as well as increasing the lifetime of the network. Currently, several types of nodeclustering algorithm exist each with different clustering attributes as explained earlier. The following section will discuss some of these techniques.
Energyefficient stabilityoriented clustering: The first design for
a heterogeneous WSN was reported by Smaragdakis et al.
(2004). This study examined the effect of a heterogeneous network with a
clusterbased clustering implementation scheme. Furthermore, it reported a simulation
of the LEACH routing protocol for heterogeneous WSNs. The LEACH routing protocol
was developed for homogeneous network implementation.
According to Smaragdakis et al. (2004), when
the node population starts to decrease, the number of cluster heads chosen per
round becomes unstable (fewer than intended) and therefore there is no guarantee
that a constant number of cluster heads (equal to nxP_{opt}) will be
selected per round per epoch. Consequently, a new clustering algorithm was proposed
to increase the stable region and thereby decrease the unstable region and improve
the feedback quality of the wireless clustered sensors in the presence of a
heterogeneous network. Some percentage of the population of sensor nodes is
equipped with additional energy resources; this is a source of heterogeneity
which may result from initialization. Hence, two types of nodes, called normal
and advanced nodes, were designed. The selection of each cluster head depends
on a threshold calculation as shown in Eq. 3 for a normal
node and Eq. 4 for an advanced node:
where, P_{nrm} and P_{adv} are the weighted probabilities for normal and advanced nodes, r is the current round identifier and G’ (G”) is the set of normal (advanced) nodes that have not become the cluster head. In the clustering process, a normal node (S_{nrm}) or an advanced node (S_{adv}) will choose a random number between zero and one. If the number selected is less than a threshold value T(S), the node becomes a CH for the current round. Otherwise, the node is not selected as a CH, but possibly in the next round it will have a chance to become the new CH. The simulation revealed that network stability was achieved and that the lifetime of the network could have been increased.
Energyefficient adaptive clustering: In an energyefficient adaptive
clustering algorithm, sensor nodes organize themselves into local clusters,
with one node in each cluster acting as the cluster head node. Certain factors
must be considered in selecting which node will become the cluster head. Careful
selection can prolong the lifetime of the Heterogeneous Wireless Sensor Network
(HWSN). Qing et al. (2006), Changmin
and Hong (2007) and Elbhiri et al. (2010)
revealed that consideration of the residual and average energy of the network
during clusterhead selection can increase the lifetime of the HWSN. Moreover,
distance (JiunJian et al., 2009) and communication
cost (Liu et al., 2011) attributes can be added
to the clustering algorithm to enhance the lifetime of the HWSN.
Qing et al. (2006) proposed a clustering algorithm
for twolevel HWSNs called Distributed Energyefficient Clustering (DEEC). The
firstlevel nodes are designated as the advanced nodes and the second level
consists of the normal nodes. The total initial energy of the HWSN is given
by Eq. 5:
where, N is the number of nodes, E_{0} is the initial energy of the normal nodes and m is the proportion of advanced nodes which possess a times more energy than normal nodes. In addition, a multilevel HWSN was considered and its total initial energy was calculated using Eq. 6:
where, a_{i} is times more energy for node i.
In the clustering algorithm sequence, the nodes were distributed uniformly and one node in each cluster was selected as the cluster head node. The selection of the cluster head was based on the initial and residual energy left in the node at each round. The nodes with high residual energy become cluster heads more often than lowerenergy ones. To organize the nodes into an HWSN, a weighted probability equation has been defined for normal and advanced nodes and is stated as Eq. 7 for a twolevel HWSN:
where, p_{opt} is the average probability of becoming a cluster head,
E_{i}(r) is the residual energy of node s_{i} in round r and
is the average energy in round r. The weighted probabilities for multilevel
HWSNs are given by Eq. 8:
The
value was estimated to make sure that the DEEC protocol can guarantee that all
nodes die at approximately the same time. Each noncluster head node sends L
bits of data per round to the cluster head.
Changmin and Hong (2007) have implemented a similar
clustering algorithm concept as a DEEC protocol. Their work showed that consideration
of the ratio between remaining energy and the average energy of the network
can extend the lifetime of the wireless network.
Incorporation of energy into clusterhead selection has been studied by Elbhiri
et al. (2009) for twolevel HWSNs. Their work was based on a DEEC
scheme with certain improvements involving the use of a threshold residual energy
value, as shown in Eq. 9. This equation is intended to ensure
that advanced and normal nodes have the same probability to become a cluster
head. As a result, clusterhead selection will be more balanced and more equitable:
where, E_{disNN} is the energy dissipated by a normal node in each round and E_{disAN} is the energy dissipated by an advanced node. Finally, the weighted probability that each node will be selected as a cluster head is given by Eq. 10:
where, c is a real positive variable which directly controls the number of cluster heads.
JiunJian et al. (2009) implemented an algorithm
similar to the SGCH algorithm but in a heterogeneous environment. In this algorithm,
every node is divided into several clusters according to the distances between
nodes. There are three types of nodes, referred to as normal, moderately advanced
and highly advanced nodes. In the clustering process, the node in each cluster
with the maximum energy remaining will be selected as the cluster head. Subsequently,
each node forwards its sensed data to the nearest cluster head. The cluster
head will then transmit the data to the sink node.
This network performs an environmental monitoring task and its sensor nodes
monitor a variety of objects. A recent study by Liu et
al. (2011) involved nodes distributed randomly with mobile or stationary
properties. They assumed that the node communication links are symmetric and
that nodes do not have any location information, but can calculate the distance
between nodes according to the signal strength received. The distance between
nodes i and j was calculated using Eq. 11:
where, α is a distanceenergy gradient with a value varying from 16 according
to the physical environment in which the sensor networks operate;
is the transmission energy for broadcasting and
is the received signal strength (received energy). In the clustering algorithm,
the node establishes a routing table of neighboring nodes based on received
data and saves all relevant information for all nodes within its communication
range. The nodes with more residual energy should be given greater opportunity
to become cluster heads and all nodes take turns at being cluster head nodes.
The threshold equation for the probability of selecting sensor nodes as a cluster
head is given here as Eq. 12:
where, α and β are coefficients, r is the round identifier, _{}
is an energy factor and
is a communication factor. The algorithm operates in an iterative fashion. After
completion of one round, a new node must be selected as the cluster head for
the next round. To reduce the energy consumed when broadcasting and the communication
cost, an energy consumption prediction mechanism for Regular Data Acquisition
(RDA) nodes has been developed. Each node will determine whether its current
residual energy is close to the residual energy predicted in the last round
using Eq. 13:
where, the value of γ is less than the amplifier energy, E_{jr,predicted} is the predicted residual energy of node j at the beginning of round r and E_{jr} is the current energy of node j.
Energyefficient multilevel clustering: Multilevel clustering is achieved
by multilevel cluster head node selection during clustering algorithm execution.
According to Hasnaoui et al. (2010) and Soni
and Katiyar (2011) use of this technique can reduce and balance energy consumption
and hence prolong the lifetime of a WSN. This can also be achieved using multihop
data transmission, in which each cluster head acts as a relay node during data
transmission (Kumar et al., 2010; Han,
2010). In addition, the network can be divided into two or more zones or
tiers to prolong its lifetime, as suggested by Li et
al. (2007a) and Mehrotra and Leong (2009).
Hasnaoui et al. (2010) carried out studies in
a twolevel HWSN of a multilevel clustering algorithm known as the Oriented
Cluster Formation (OCFLE) algorithm. OCFLE is an oriented cluster formation
algorithm based on the position of the CHs in the sensing area and on the minimum
distance between a noncluster head node and the sink node using a suitable
intermediate CH.
In the clustering process, the cluster head node was selected based on the ratio between the residual energy of each node and the average energy of the network. A twolevel hierarchical design was introduced in which suitable intermediate CHs for data transmission were developed. The CH with the most energy, called MaxCH, is the most likely to be chosen as the intermediate CH used to transmit aggregate data to the Base Station (BS). MaxCH collects all data coming from all CHs, compresses them into a single signal and sends them directly to the base station. The multilevel clustering network concept is illustrated in Fig. 4.
Recently, Soni and Katiyar (2011) have developed a
clustering algorithm for twolevel HWSNs which introduces the concept of Minimum
Reachability Power (MRP). This concept takes into consideration the transmission
range among the nodes in the network. Eq. 14 expresses the
probability of cluster head node selection during the clustering process:
where,
is the probability of becoming CH for a normal node and
the probability of becoming CH for an advanced node, φ is a parameter which
determines the weighting of energy and distance factors,
is the optimal number of CHs for level 1, E_{u}(t) is the energy of
the nodes, γ is the energy multiplier for advanced nodes compared to normal
nodes, Σ_{v∈S} E_{v}(t) is the total remaining energy
of the network and P_{u} and P_{v} are the MRP values. Eq.
15 shows the probability calculation for level2 clusterhead selection:
where, N_{1} is the cardinality of the active set.
After a levelT clustering topology has been formed, the regular nodes start
transmitting their sensed data to their CHs. LevelT CHs aggregate the sensed
data and send them to level(T1) CHs and so forth. Finally, all level1 CHs
transmit the aggregated data to the sink node.
Kumar et al. (2010) extended the clustering algorithm
to include intercluster transmission considerations. Their research is an extension
of their previous work on the EEHC algorithm. A new mechanism was added to the
clustering algorithm, as shown in Fig. 5. The diagram shows
that the sensed data were transmitted to the sink node using multihop data
transmission. The path metrics and their route costs were formed into an array
matrix called an “adjacency matrix.” In each cluster, the cluster
head receives the messages and then aggregates the received messages with its
own message. The aggregated message is then forwarded to the nexthop receiver
(sink or CH), based on the routing information calculated using the optimal
path approach.
Han (2010) examined multiplehop routing in a clustering
algorithm for HWSNs and proposed a threelevel heterogeneous network. The nodes
organize themselves into local clusters, with one node acting as the cluster
head in each cluster.
The node with the most remaining energy will be selected as the cluster head.
The CH node sets up a TDMA schedule and transmits it to the nodes in the cluster.
Noncluster head nodes will join a CH based on the distance calculation given
in Eq. 16:
where, d is the distance and E_{remaining} the remaining energy of the clusterhead candidate. Each CH finds its own optimum path to the sink node using a minimum spanning tree algorithm. This technique can be used for multiplehop data transmission which can achieve short communication distances between the cluster heads and the sink node.

Fig. 6: 
Schematic diagram of differentsized zones 
Li et al. (2007b) investigated a zonebased
clustering algorithm for twolevel HWSNs called zonebased REECR (ZREECR). Their
work was based on the Residual Energy and Energy Consumption Rate (REECR) routing
protocol. In the clustering process, the cluster head nodes were evenly spread
throughout the network which can minimize the distance over which the noncluster
head nodes need to send their data. The algorithm uses a zonebased routing
protocol which divides the network field into several zones of different sizes
depending on their distance and orientation from the sink node. Figure
6 shows an example of a twolayer zone of the network.
Cluster head nodes closer to the sink node have smaller zones because they
consume less energy for intracluster data processing and can preserve more
energy for intercluster relay traffic. Consequently, the choice of the cluster
head in each zone is based on residual energy and on energy consumption rate,
as shown in Eq. 17:
where, α and β are weighting coefficients, E_{i}(t) is the
current residual energy of node i and V_{i}(t) is the energy consumption
rate of node i. Hence, all the data collected in the cluster head will be forwarded
to the base station over multiple hops, in which every cluster head chooses
to forward its data to the closest cluster head in the direction of the base
station. As a result, this algorithm is approximate, but not more energyefficient
than REECR. Mehrotra and Leong (2009) conducted a simulation
of a twolevel HWSN in various network tiers, as shown in Fig.
7. The lowest tier is closest to the base station. Every sensor node belongs
to one of these tiers.
This kind of network is guaranteed to have two properties: first, to maintain a short distance between transmitter and receiver and second, to choose relay nodes on the basis, not only of each node’s distance from the transmitter, but also of its residual energy. In the clustering process, this algorithm uses a nodebynode adaptive approach which can use either direct transmission or the multihop scheme depending on the residual energy limitations of each node. The communication process in the network is governed by two sets of rules. The first set of rules applies when the residual energy of a node is greater than 30% of its initial energy while the second set of rules applies when the residual energy of a node is less than 30% of its initial energy. This technique can increase the efficiency of the algorithm and help to prolong its period of stability.
All the algorithms discussed in this section are summarized in Table 2. Table 3 then classifies these algorithms based on the attributes discussed. Finally, Table 4 presents a comparison of the clusterhead node selection process used for the various algorithms.
PERFORMANCE EVALUATION METRICS
Some of the parameters that can be used to evaluate the performance of nodeclustering algorithms are:
Average energy dissipated: The average energy usage by a sensor node in the network over time for various purposes such as transmitting, receiving and aggregating data.
Table 2: 
Comparison of various types of energyefficient nodeclustering
algorithm 

Table 3: 
Classification of energyefficient nodeclustering algorithms
based on clustering attributes 

Network lifetime: The time interval from initial node activation until network death.
Number of cluster heads per round: The number of cluster heads chosen in each round which represents the number of nodes which will send data directly to the sink node. Comparison of cluster head selection approaches is shown in Table 4.
Total number of nodes alive: The number of nodes that remain alive in each round.
Number of data messages/throughput: The number of data messages received by the sink node. If the volume of data messages received is too big for the sink node, it will make the sink node die and energy consumption will decrease.
Table 4: 
Comparison of clusterhead selection approaches 

CONCLUSIONS AND FUTURE TRENDS
A survey of the current stateoftheart in energy efficiency for various types of nodeclustering algorithms has been discussed. A number of suggested future research directions are described below.
Hotspot problem: Multihop data communication is a good technique for achieving energy efficiency; however, it has a “hotspot” problem which can force a number of nodes in a certain area in the network to die very early.
Coverage: Good coverage ensures that all the nodes in the network provide the data required by the sink.
Optimization of energy efficiency: The energy efficiency of a nodeclustering algorithm can be optimized using various optimization techniques such as genetic algorithms and antcolony algorithms.