Performance Comparison of Various Clustering Techniques in Wireless Sensor Networks

This paper presents the comparison of various Wireless Sensor Network (WSN) clustering protocols like LEACHCentralized, KMeans based clustering, Fuzzy C-Means clustering and Harmony Search Algorithm based clustering. The protocols have been compared with respect to network lifetime, energy consumption and effectiveness of clustering. It was observed that HSA based method of clustering outperformed LEACH-C, K-Means and FCM based protocols in almost every performance measurement aspects. However, few shortfalls of HSA method were also identified. All these protocols only consider intra cluster distance of members while clustering. Due to that the data delivery of these protocols are not uniform and clusters are not covering the deployment field in uniform way. The clustering problem was also observed for these protocols, wherein large variation was found in the cluster sizes during a particular round of the clustering algorithms Keywords LEACH-C, FCM, Harmony Search, Clustering. 1 Introduction Communication Technology foster the fast growing world with quick and reliable data transfer and information sharing. With the advent of various generations of networking technologies, applications have been developed for automatic data acquisition and information compilation and computation. Sensor networks is one such application where hundreds to thousands of sensor nodes are deployed for gathering information related to environment conditions, performing necessary simple computations and sharing the information with peer nodes or with an external Base Station. A sensor network can be created by manual as well as random deployment. This results into a fault tolerant and maintenance free network for data gathering. Since these sensor networks employ nodes working on battery, it is also important to conserve energy of the nodes so that the life time of network is enhanced. The research of sensor network primarily focuses on reducing the energy drain of the nodes, thereby increasing network life. Data sensing and reporting in sensor networks is dependent on the application and time criticality of data. So, a sensor network can be categorized into time-driven or event driven networks. Time driven network is suitable for applications that require periodic data monitoring, wherein nodes periodically switch on their sensors and transmitters, sense the environment and perform data transmission at constant periodic time intervals. Hence, through this type of network, snapshot of relevant attributes are provided at regular intervals. In the event driven network, a monitored external event triggers the data acquisition and information is transmitted to the base station or peers. There are many challenging factors which highly influence the design of WSN routing protocols [1]. Deployment of sensor network is random and hence ad hoc. This mandates the system to be able to cope up with the distribution of nodes and establish connections between nodes. Due to limited energy supply, computation and transmission as per the application requirements in a wireless environment with reliability is highly difficult. Limited computation capabilities of sensor nodes does not allow porting of sophisticated network protocols, as is the case with wired networks wherein the nodes are rich in terms of resources. Since communication range of sensor nodes is limited, route will generally consist of multiple wireless hops for communicating with peers or Base Station. A sensor node consists of many hardware components, integrated in less than a cubic centimetre. Such components consume extremely low power and operate in an unsupervised mode. They must also adapt to the environment of sensor network precisely. Collisions generally adds more trouble in to the network like retransmissions, which in turn triggers latency problem and increases energy usage as well. The overhead of control packet in this scenario linearly related to the density of nodes. At times, data is time-critical within a particular duration of time. If failed to deliver during this deadline, the data becomes useless. For time-constrained applications, bounded latency is yet another constraint. To deal with these constraints, sensor networks need highly customized routing protocols. Routing protocols specifically designed for sensor networks highly depends on application requirements. Major categories as mentioned in [2][3], are Flat which are data centric, Hierarchical and Location based. These protocols are further classified into multipath based, query based, negotiation based, QoS based and coherence based IJCSC VOLUME 5 • NUMBER 2 JULY-SEPT 2014 PP. 116-123 ISSN-0973-7391 117 protocols. However, some hybrid protocols exist that fit under more than one categories. D J Dechene et al [4] explored various routing algorithms based on clustering for WSNs and classified them into Heuristic, Weighted Schemes, Hierarchical Schemes and Grid based approaches. Specifically they examine performance of these algorithms in terms of the power and quality aspects of these schemes. G Santosh Kumar et al [5] discussed four routing protocols for wireless sensor networks viz. Flooding, Gossiping, Gradient Based Routing (GBR) and LEACH. They simulated these protocols using TinyOS. As sensor network deployments for critical applications are generally large scale in nature, specific routing techniques are required to handle such large networks. Scalability and efficiency of communication is also achieved using Hierarchical routing technique. Here, higher energy nodes can be used to process and send information, while low energy nodes can be used to perform sensing in the proximity of target. Many such routing protocols have been proposed in the literature such as LEACH [6], LEACH-C[7], KMeans[8][9], FCM[11][13][14][15], HSA[16][17][18]. LEACH is considered as the most popular routing protocol that use cluster based routing in order to minimize the energy consumption. A variant of LEACH, LEACH-C is a centralized version of LEACH which exploits base station for clustering purpose. In this paper, we analyse the performance of various centralized clustering protocols. In section II, we discuss various clustering techniques in detail. Section III present the simulation results and analysis of these methods. Section IV and V conclude the paper. 2 Clustering Techniques Clustering approaches have become an emerging technology for building scalable, robust and energy efficient applications in order to ensure efficient network operations. Conserving energy resources in such networks is considered as the most important research avenue. As per the review of literature, many contributions have been made for minimizing energy usage during communication. Data is routed form all nodes to a sink node using the process called convegecast. Moreover in these applications it is observed that, energy is consumed more in data transmissions instead of data processing. Hence, an energy efficient technique called data aggregation is used which collects local data at intermediate nodes and forwards the aggregated data. Data aggregation methods are generally synchronized with the packet flow in the network. The protocols have been designed by researchers which allows innetwork processing along-with routing of packets. These type of protocols fall in to two different categories, namely tree based aggregation and cluster based aggregation. Tree based aggregation may suffer from latency when number of nodes are large. To control the latency, clustering techniques may be applied. Every cluster reports aggregated data to the base station. Within a network, all sensor nodes group themselves in to several regions using short messages in cluster based data aggregation technique. Here, at any instance one node acts as a cluster-head and sends short messages to other remaining nodes. Depending upon the signal strength the sensors decide to join the groups. LEACH [6] is very popular protocol used by researchers in hierarchical routing related work. LEACH attempts to minimize energy dissipation. It is based on a simple clustering mechanism by which energy can be conserved since cluster heads are selected for data transmission instead of other nodes. LEACH operates iteratively in rounds till all nodes are not expired. Each round has a setup-phase and steady-state-phase. In setup phase clusters are organized and in steady-state phase data is transferred to sink. To reduce overhead setup phase is shorter compared to steadystate phase. Tyagi and Kumar [10] present detailed review of LEACH related protocols. The literature review suggest that LEACH protocol suffers from several problems, one of them is non-uniform clustering. Better clusters may be produced using centrally controlled algorithm called LEACH Centralized (LEACH-C) [7]. It has similar phases like LEACH. The steady-state phase of LEACH-C is exactly similar to LEACH. During set-up phase of LEACH-C, all nodes send their geographic details and present energy level to base station. BS calculates the average network energy, and assigns the role of cluster-heads (CH) to the members who are having energy above the threshold for the current round. BS finds clusters using the simulated annealing algorithm [12]. LEACHC controls energy usage of non-cluster head nodes, by creating clusters wherein the cluster-head and members are as close as possible. The algorithm assigns the members to clusters in an optimized way, by minimizing sum total of squared distances between all member nodes and the candidate cluster-head. After the cluster-heads and associated cluster members are finalized, BS broadcasts message containing cluster-head ID for each node. All receiving nodes match their own ID with the received cluster-head ID, to determine whether they are cluster-heads or not. If a node is not designated as cluster-head, it determines its TDMA slot for data transmission and changes the operating mode to sleep state until the desired TDMA slot is available to transmit data. The energy model used is as follows. Energy required to transmit L bits at a distance d, (1) IJCSC VOLUME 5 • NUMBER 2 JULY-SEPT 2014 PP. 116-123 ISSN-0973-7391 118 Where d0 is the deciding factor whether to use free space propagation model or multi-path radio propagation model. fsand mp are the amplification components depending on the propagation model in use. Energy required to receive L bits, In [9], authors have proposed K-Means based clustering where, they minimize the objective function, ! " #! $ %& '!%& (2) Here ! " #! is a chosen distance measure between a data point Xand the cluster centreCj. It is an indicator of the distance of the n data points from their respective cluster centres. K-Means Algorithm [9] creates ’k’ centroids randomly and then computes the Euclidian Distance of each node with every centroid. It iteratively computes the location of centroid in each cluster and looks for change in the position with respect to previous iteration. If change is found in the centroid position, the algorithm continues, otherwise the clusters are considered final and the algorithm converges.In [8], authors have proposed BPK-Means clustering algorithm. They refer the algorithm as balanced parallel K-Means algorithm. Authors balance the cluster size by keeping the number of members exactly same in every cluster.In [11] and [15], authors have applied Fuzzy C-Means clustering algorithm for centralized clustering. The objective function to be optimized is, ( ) ! * " #!* +!%& $ %& ,-./. 0 1 2 3(3) where ) ! & 456789:5 *6789;*< = >8? 9;@? (4) The norm, * " #!*measures the similarity (or closeness) of the data point Xi to the centre vector Cj of cluster j. Cluster centers are computed with, #! A7: >B 7 C7@? A7: > C7@? (5) The degree of membership for data point i to cluster j is initialised with a random value) !, D 1 ) ! 1 0, such that ) ! +! 0BInitially the degree of membership is assigned randomly. As the algorithm converges, clusters are evolved with termination criteria. The termination conditions are the least overlapping membership of members with neighboring clusters and the fuzziness coefficient of the membership. The FCM algorithm terminates when the desired accuracy is achieved for the degree of membership, till then the algorithm iteratively executes and compares the previous and current degree of membership. The largest difference between the current and previous degree of membership for all data points with all clusters is the termination acts as termination criteria here. Once the clusters are created with the centroids, the members near the centroid are labelled as cluster-head depending on members’ available energy. Harmony Search Algorithm was initially proposed by Geem and Kim[19]. Later applied for clustering process in WSN by Hoang, Kumar and Panda[16]. Authors find the best clusters using intra cluster distance and residual average energy. They recommend nodes with higher residual energy to become cluster head along-with consideration of intra cluster distance of cluster members from cluster heads. They define Harmony Memory as matrix HM (eq. 6) having randomly selected member elements in every row. Each vector in the matrix HM indicates that elements in the row may be cluster head after the algorithm converges. These vectors are partially altered or replaced altogether with harmony improvisation logic. EF G H&& H& I H&J H & H I H J I HKL M & I HKL M II I HKL M J N O & I KL MP(6) Once the HM is derived, a new harmony memory is devised. The new Harmony from HM is derived using Harmony Memory Consideration Rate (HMCR). A new harmony may be directly selected from the candidates present in the IJCSC VOLUME 5 • NUMBER 2 JULY-SEPT 2014 PP. 116-123 ISSN-0973-7391 119 HM with probability HMCR or they may be generated fresh from the pool of all the sensor nodes which are alive using the probability 1 HMCR. H!Q R H!Q S EF T ,UVW/XYZYU[UV\ EF #] H!Q S H ̂ _` `^a ,UVW/XYZYU[UV\ 0 " EF #] (7) If the new harmony has been derived from HM itself then it is further fine-tuned with Pitch Adjustment Rate (PAR). With probability PAR, member elements in the new harmony may be mutated with other members in the same position in the Harmony Memory having residual energy higher than the average energy. Otherwise the new harmony will be left as it is with the probability 1-PAR. Pitch Adjustment for selected Vj ` is, H!Q R H!_ S EF ,UVW/XYZYU[UV\ bc] H!Q ,UVW/XYZYU[UV\ 0 " bc] (8) The goal is to minimize the objective function: def!>7g h i d& 0 " h i d (9) where d& 2Zj!S & J k ` _e` 7 +K: lgmno7S9: +: p (10) The function f1 finds the maximum of the average member distance with the cluster head of every cluster. d q7ros lgmno7S9: q9t: ros u J!%& (11) The function f2 finds the effect of residual energy of CH on to the sum of total energy of every member in every cluster. The objective function is designed in such a way that this term’s contribution is minimized, thereby ensuring that members with higher residual energy become cluster-heads. 3 Simulation Results and Analysis Protocols discussed in previous section were simulated in NS-2 with C++ and TCL. Table 1 shows the network specific simulation parameters used. Table 2 shows the energy dissipation parameters used during simulation. The simulations were performed for different network parameters settings. The deployment was varied from 50 to 200 meters. The number of nodes were changed from 50 to 200 nodes. The base station location was also changed between center position and 50,175 in the field. Figure 1 shows the network performance in terms of number of nodes alive during the simulation. It may be observed that Harmony Search Algorithm preforms better compared to other protocols, as it looks for the best possible clustering using evolutionary technique. Table 1: Simulation Parameters Network Parameter Value Node distribution (0,0) to (200,200) BS location Center and (50, 175) No. of Nodes 50,100,150,200 Initial Node Energy 2J Simulation Time 3600s Desired No. of cluster-heads 5% (Optimal) Bandwidth of the channel 1 Mbps Packet header size 25 Bytes Message size 500 Bytes IJCSC VOLUME 5 • NUMBER 2 JULY Operation Energy consumed for transmitting and receiving Energy consumed by amplifier to transmit at shorter distance Energy consumed by amplifier to transmit at longer distance Energy consumed while data aggregation Figure 1: Alive nodes for LCO, KMeans, FCM and HSA Figure 2 shows the data reported by various protocols during simulation to the base station via cluster heads. The data reporting of HSA and FCM based methods are almost similar. But carefully examining Figure 1 shows that number of alive nodes quickly falls in FCM whereas, in HSA it is gradually falling, but it is still performing better compared to other protocols when number of alive nodes is the comparison criteria. Figure 3 shows the clustering done by KMeans algorithm for 100 nodes deployment. It wa round that which nodes are elected as cluster that one cluster-head is assigned only 2 member nodes while another cluster clustering method is not distributing members uniformly across all clusters. Similar problem was also observed for FCM based clustering. It is presented with Figure 4. Here also, one cluster whereas the other cluster-head is assigned more than 30 nodes, out of total 100 nodes. Figure 2: Data Delivered for LCO, KMeans, FCM and HSA -SEPT 2014 PP. 116-123 ISSN-0973-7391 Table 2: Radio parameters Symbol Energy dissipated Eelec 50nJ/bit (dtoBS d0) fs 10pJ/ (dtoBS d0) mp 0.0013pJ EDA 5nJ/bit 200 nodes deployment s captured at a particular -heads and how many members are assigned to them. It may be seen -head is assigned close to 50 nodes. So -head is assigned less than 10 nodes 150 nodes deployment