A genetic fuzzy system based optimized zone based energy efficient routing protocol for mobile sensor networks (OZEEP)

Accepted Manuscript Title: A genetic fuzzy system based optimized zone based energy efficient routing protocol for mobile sensor networks (OZEEP) Author: Juhi R. Srivastava T.S.B. Sudarshan PII: DOI: Reference:

S1568-4946(15)00597-9 http://dx.doi.org/doi:10.1016/j.asoc.2015.09.025 ASOC 3211

To appear in:

Applied Soft Computing

Received date: Revised date: Accepted date:

1-12-2014 12-9-2015 14-9-2015

Please cite this article as: J.R. Srivastava, T.S.B. Sudarshan, A genetic fuzzy system based optimized zone based energy efficient routing protocol for mobile sensor networks (OZEEP), Applied Soft Computing Journal (2015), http://dx.doi.org/10.1016/j.asoc.2015.09.025 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Ac ce

pt

ed

M

an

us

cr

ip t

Highlights  Genetic fuzzy system based optimized clustering over existing routing protocol ZEEP  Cluster head selection criterion are energy, distance, node density and mobility  Fuzzy module is used to nominate cluster head candidates in the first phase  Genetic algorithm further optimizes the selection from the nominated candidates  Achieved balanced clustering for enhanced network lifetime and minimum packet loss

Page 1 of 36

A genetic fuzzy system based optimized zone based energy efficient routing protocol for mobile sensor networks (OZEEP) 1

Juhi R. Srivastava*, 2T.S.B. Sudarshan

Department of Computer Science and Engineering, Amrita VishwaVidyapeetham University, Amrita School of Engineering, Bangalore, India

* 1 2 Corresponding author +91 9886396059 [email protected], [email protected]

ip t

________________________________________________________________________________________ Abstract

M

an

us

cr

Wireless sensor networks have become increasingly popular because of their ability to cater to multifaceted applications without much human intervention. However, because of their distributed deployment, these networks face certain challenges, namely, network coverage, continuous connectivity and bandwidth utilization. All of these correlated issues impact the network performance because they define the energy consumption model of the network and have therefore become a crucial subject of study. Well-managed energy usage of nodes can lead to an extended network lifetime. One way to achieve this is through clustering. Clustering of nodes minimizes the amount of data transmission, routing delay and redundant data in the network, thereby conserving network energy. In addition to these advantages, clustering also makes the network scalable for real world applications. However, clustering algorithms require careful planning and design so that balanced and uniformly distributed clusters are created in a way that the network lifetime is enhanced. In this work, we extend our previous algorithm, titled the Zone-based Energy Efficient Routing Protocol, for Mobile Sensor Networks (ZEEP). The algorithm we propose optimizes the clustering and cluster head selection of ZEEP by using a Genetic Fuzzy System. The two-step clustering process of our algorithm uses a Fuzzy Inference System in the first step to select optimal nodes that can be a cluster head based on parameters such as energy, distance, density and mobility. In the second step, we use a genetic algorithm to make a final choice of cluster heads from the nominated candidates proposed by the fuzzy system so that the optimal solution generated is a uniformly distributed balanced set of clusters that aim at an enhanced network lifetime. We also study the impact and dominance of mobility with regard to the variables. However, before we arrived at a GFS-based solution, we also studied fuzzy-based clustering using different membership functions, and we present our understanding on the same. Simulations were carried out in MATLAB and ns2. The results obtained are compared with ZEEP. Keywords: mobile; sensor; routing; mobility; clustering; genetic algorithm; fuzzy logic, fuzzy inference engine

perform transmission of the sensed data, routing and sleeping. Figure 1 shows these phases of operation.

pt

1. Introduction

ed

________________________________________________________________________________________

Ac ce

Demand for ‘Ambient Intelligent’ devices has become very popular in our daily lives. Such devices are capable of sensing and controlling environmental or physical parameters such as temperature, humidity, pressure, and movement and allow remote human interactions with applications where direct intervention of humans is difficult or is impossible. Networking such embedded systems, called sensors, create a new and versatile network known as Wireless Sensor Networks. Due to the low cost, easy integration and mass-scale production of MEMS technology, wireless sensor networks have garnered great deal of interest in a variety of applications such as the design of intelligent buildings, disaster relief, military, facility management, medicine and health care, and machine surveillance. However, such varied applications require very careful planning, design and implementation of such sensor networks so that they achieve complete coverage and connectivity for accurate detection or monitoring of events. After the sensor network deployment covering the physical space under observation, each node in a sensor network goes through multiple phases of operation. Every node starts by sensing a particular event, listening for the communication channel to be free to

Fig.1 Various phases of operation of a sensor node The activities that are primarily responsible for energy consumption in the sensor networks are sensing events and data routing. Energy consumption

Page 2 of 36

ip t

Fig.3 Energy analysis of direct transmission

cr

Fig.4 Energy analysis of multi-hop transmission

us

The biggest challenge associated with these nodes is that they are extremely energy constrained and have variable channel capacity and limited bandwidth availability to perform data routing. Such networks also have redundant information and, because of best path routing, have cluttered loads on a single or a few paths compared to other available routes, giving rise to an increased unbalanced load and data losses in the network. In such a highly resource-constrained scenario, delivering real time data and maintaining real time communication along with quality of service are very stimulating research topics, in particular when such networks exist in harsh environments with heterogeneous nodes trying to address mobility, dynamic network topology, node failures and communication failures. It is, therefore, very important for such networks to be self-configurable in nature and operate in an adhoc manner. Figure 2 shows the collaborative processing performed by sensor nodes. It has been observed that in reality, real world problems and applications require a network with variable size. In such networks, the number of nodes deployed grows generously, leading to an increased network size. Therefore, scalability demands a network architecture that can meet the reduced energy consumption requirement in addition to being fault tolerant, allowing load balancing, and performing efficient data routing. Clustering or grouping of sensors is a well-known two-tier architecture that has proved [1] to be a good solution for reducing network energy consumption as well as meeting the scalability requirements. Clustering of nodes allows efficient data aggregation and selection of optimal paths toward the sink. They help reduce generation of duplicate packets in the network and further save the network from message overheads for route creation every time a node has to send data. Clustering, which is a hierarchical model of data transmission, restricts participation of the nodes in dissipation of information and allows multi-hop data routing only between the cluster heads (CH). Hence, with a reduced number of nodes performing data transmission, the overall

ed

M

an

by nodes varies during the sensing of events depending on whether the application requires continuous or periodic sensing of events. Accordingly, diverse studies have been performed in the literature that focus on the sleeping and sensing patterns of the nodes in the data link layer, in particular the Medium Access Control or MAC layer. In the data routing phase, all of the sensors that are active transmit their sensed information to the sink or the base station. It is observed that in direct communication, nodes that are located close to the sink or the base station spend less energy during transmission than the nodes that are placed at greater distances from the base station or the sink node. This unequal energy consumption leads to a multi-hop routing choice and, therefore, a many-to-one traffic pattern in the network. Figure 2 shows a diagrammatic representation of the data packet arriving from multiple sources.

pt

Fig.2 Collaborative processing between multiple data sources and the base station

Ac ce

An energy analysis of routing data using direct transmission reveals that as the strength of the signal reduces by square of the distance (di) therefore, the nodes located far away from the base station (BS) have to expend more energy in transmitting their data to the sink. The amount of energy used in direct transmission can be modeled as ƐampK (3d1+d2)2, where Ɛamp is energy used by amplifier and K is constant. It is also evident that the nodes farthest from the BS die out quickly giving rise to network partitions and low network lifetime. However, in multi hop communication, the data is carried to the sink via multiple hops through closely located nodes or neighbor nodes. The amount of energy used for data transmission in multi hop communication is given by ƐampK(3d12+d22) and is much is less due to smaller transmission distances involved when compared to direct transmission. (Figures 3 and 4 show the energy analysis).

Page 3 of 36

nodes in the network have the same power level at any point of time. Hence it becomes extremely necessary to rotate the job of CH among the other nodes so that they can share the role of CH for balanced energy dissipation in the network. This means that to rotate the role of CH we cannot run the clustering algorithm only once but will need to execute it repeatedly. The new CH which is selected at each repetition will have better values for energy along with other network criteria that have been defined for the CH selection in the clustering algorithm. The node which was a CH in the previous round will not participate in the CH selection as it will have lower remaining energy at the end of the previous round when compared to other nodes who were non CH in those rounds. This concept of selecting a new CH based on the changing network parameters like node’s remaining energy, density around the node, node’s distance to base station and node’s mobility with an aim to have complete network coverage and evenly distributed network’s energy dissipation is called dynamic clustering. In this method once clustering is completed (clusters are formed and respective cluster heads are selected) no CH is changed until data transmission in the steady state phase is over. Only when re-clustering is initiated, a new CH is selected again based on the current network dynamics. In addition, it is important that when clustering is performed, all sensor nodes should belong to at least one cluster so that network partition does not happen. However, another major concern that originates while designing clustering algorithm is that the number of overlaps between nodes and clusters should be kept to a minimum value. It becomes extremely necessary to devise efficient clustering algorithms that do not only rotate the role of cluster heads in the network efficiently but also choose optimal routing paths so that the overall energy consumption of the network is reduced considerably to improve the network lifetime. In the literature, many algorithms have been proposed to address these issues in the best possible way. Over the years, optimization of classical routing protocols has been done with the aim of extending network lifetime. However, WSN challenges require an intelligent and flexible technique to process the ambiguity and uncertainty related to node behavior and data in the complex network environment. This has prompted many researchers to study and embed various soft computing techniques in their design of optimal algorithms. Soft Computing is the blend of methodologies that were designed to sculpt and facilitate answers to real world problems, which are not modeled or are too difficult to model mathematically. These problems are typically associated with fuzzy, complex, and dynamical systems with uncertain parameters. The uncertainty associated with the network state information is unavoidable in terms of both fuzziness and randomness. The main techniques in soft computing are Bayesian statistics, evolutionary

Ac ce

pt

ed

Fig.5 Process flow in a Wireless Sensor Network

M

an

us

cr

ip t

energy consumption of the network is reduced significantly, which results in an extended network lifetime. Figures 5 and 6 represent this idea.

Fig.6 Cluster architecture used in a Wireless Sensor Network As we can see an important component of clustering is a cluster head (CH) which performs many activities like organize the medium access within the cluster, participate in routing decisions. Cluster heads are also natural places for data aggregation and data compression and hence it is evident that the cluster head of each cluster expends more energy than any other node in the network. It is often desirable that all

Page 4 of 36

Ac ce

pt

ed

ip t

cr

us

M

In this work, a genetic fuzzy (GFS)-based method to optimize the clustering and routing model of the Zonebased Energy Efficient Routing Protocol for Mobile Sensor Networks (ZEEP) [9] is proposed. ZEEP is a two tier routing protocol designed explicitly for wireless sensor networks. In this work, we optimize the cluster head selection mechanism of ZEEP by utilizing the benefits of the fuzzy logic system and genetic algorithm. Fuzzy logic or the fuzzy inference engine is used to help nodes decide the caliber to be elected as a CH based on parameters such as energy, distance between nodes, node density and mobility levels. We further use the evolutionary search algorithm, the genetic algorithm, on the output of the fuzzy inference system and select the best nodes as cluster heads so that the energy consumption of the network is as minimal as possible. A detailed study and understanding of WSNs made us converge on the following input variables for the fuzzy module embedded in the nodes. We briefly describe our learning process to help orient readers to understand the work described in this paper. Concerns associated with mobile networks and those of stationary networks differ significantly. For stationary networks, parameters such as distance between nodes and sinks do not vary once the network is deployed, and parameters such as energy, node density and transmission energy vary in a predictable manner, depending on the application. Hence, clusters, once formed, are not disturbed until re-clustering is performed. The scenario for mobile networks is quite different. Because of mobility and, consequently, the different speeds of nodes, the distance between nodes varies continuously. The node density is also variable here as the number of nodes surrounding the mobile node changes due to the node’s movement. Consequently, the variation in the transmission energy of the nodes also becomes significant. All of these factors impact the energy consumption of the nodes and ultimately lead to unpredictable network behavior. Thus, a good cluster configuration is disrupted more often, leading to data losses or packet drops and also intensified energy consumption by the network because of the reconfiguration of the zones and paths in addition to managing route configuration overheads. This

requires re-clustering of nodes as early as possible so that the values of clustering parameters do not change significantly and also data losses do not happen. Therefore, mobility is a vital decisive factor selected in our proposed clustering based routing algorithm OZEEP. In both cases of stationary and mobile networks, it is important that the selection of CHs is appropriately done so that CHs are optimally distributed in the network; in other words, no two CHs should be very far apart or very close to each other, they should not overlap and all nodes should belong to at least one cluster. This requirement is important for reducing the energy consumption of the deployed network. OZEEP aims at creating such stable clusters, which cannot be modeled by any classic mathematical methods. OZEEP can start with an initial set of cluster configurations and can also run on network with randomly deployed nodes. Our proposed algorithm takes into account both stationary as well as mobile nodes. Each node in OZEEP has a fuzzy module embedded in it that helps the node to judge the caliber to be chosen as a CH depending on its remaining energy levels, node density, distance and mobility factor. The number of candidates thus selected by the fuzzy inference system is a reduced number or a subset of the total number of nodes deployed. This information is then sent to a SINK or a base station, which uses the genetic algorithm to make a final selection of nodes that can actually be the CH. An optimal number of nodes to be a CH are important to ensure an even distribution of CHs in the network for load balancing and fault tolerance. In the proposed work, we have tried to use a genetic fuzzy system to select nodes as cluster heads and also decide on the optimal number of cluster heads for the network to achieve optimal load distribution, fault tolerance, efficient data gathering and dissemination. The outputs of the fuzzy system and genetic algorithm obtained from MATLAB tool boxes were fed into ns2, and our optimized algorithm OZEEP was compared with ZEEP on packet loss and energy consumption parameters. However, before we arrived at a GFSbased solution, we also studied fuzzy module in detail and performed clustering using the fuzzy system alone. We performed experiments using different membership functions and observed their behavior with respect to energy, distance, density and mobility of the node. We realized that the choice and shape of membership function selected are important to achieve optimal output, and therefore, we selected a triangular membership function as opposed to a generalized bell-shaped curve or trapezoidal membership function. Nevertheless, the foundation of the whole study is targeted to recognize the relationship of a node’s mobility with system variables such as distance, energy and node density to correctly understand the impact of mobility on the network behavior and the system performance.

an

computing or genetic algorithms, artificial neural networks, fuzzy logic, particle swarm optimization, simulated annealing, and ant colony optimization. Each technique can be used separately, but a powerful advantage of soft computing is the complementary nature of the techniques. Used together, they can produce solutions to problems that are too complex or inherently noisy to address with conventional mathematical methods. These features have made soft computing usage spread to various applications in communications, aerospace systems, electric power systems, industrial and manufacturing automation, robotics, and transportation [16].

Page 5 of 36

ip t

cr

Ac ce

pt

ed

M

The main task of clustering is to select a cluster head whose selection has a significant impact on the network performance. The clustering process has multiple stages of selecting a CH. The process starts with deciding the clustering architecture, that is, whether to choose a centralized control or a distributed approach. The next step is proceeding toward setting objectives depending on the application requirements. The whole clustering process goes through different phases of characterizing network goals and implementing the goals based on properties of CH requirements. Many clustering algorithms have been proposed in the literature for WSNs. They have been classified [1], [13] according to the network architecture such as network dynamics, node deployment capabilities and in-network processing, clustering objectives such as load balancing, fault tolerance and network connectivity, cluster properties such as inter- and intra-cluster connectivity, cluster counts, stability, cluster head capabilities, mobility, and node role. Figure 7 gives a broader view adopted for selecting a cluster head in our work presented here. Our design incorporates a distributed architecture with an aim to achieve maximum network lifetime. The design is intended for a mobile environment and can be extended to large networks, making it scalable and fault tolerant in nature. We use an attribute-based

us

2. Related works

cluster head selection technique where we consider distance, node density, mobility and energy parameters of nodes to be selected as CHs. Even though most algorithms are periodic in nature, in some cases where re-clustering is trigger based (for example the rate of degradation of node’s energy is a trigger criteria), then flooding of messages which have state information of nodes included, is used. The nodes which receive and forward the messages for the first time become the CH. They stamp the message with their IDs so that the rest of the receiving nodes are able to tell apart whether the messages they received are from a CH or an ordinary node. The node which receives stamped messages from multiple CHs understands that they are common points for these CHs and therefore become gateways. This state information helps nodes to realize their neighbors and their respective CHs. The steady state phase starts once flooding terminates. This technique of selecting a CH on fly by flooding of messages and not considering network dynamics and current state of the node which becomes a CH is called as passive clustering. The whole process of CH selection in our proposed algorithm OZEEP is dynamic in nature as every time a CH is selected, it is based on its current status in the network depending on the abovementioned attributes. Looking through the disadvantages of using predefined algorithms or a random or probabilistic approach for CH selection, we have adopted a deterministic two-tier approach for CH selection. Deterministic approaches take into account the condition of a node before proceeding to CH selection. For example, the remaining energy levels or mobility behavior of a node is observed before the algorithm makes a choice. This helps for a more realistic and fair CH selection in the network, which further dominates the increments in the network energy standards.

an

The work is presented as follows: Section 2 describes the related works to call attention to the motivation for proposing our new protocol OZEEP. Section 3 gives the system and energy model, and Sections 4-6contain the details of OZEEP operation. Section 7 presents the simulation environment, and in Section 8, we present the results and discussions. In Sections 9and 10, we give our conclusions and future work, respectively.

Page 6 of 36

ip t cr us an

M

Fig.7 Clustering modules featured in our proposed algorithm OZEEP

Ac ce

pt

ed

Research in clustering protocols for WSNs has shown that soft computing techniques have proved to be an inspiration to address uncertainty, approximation, partial truth and ambiguity associated with the sensor network environment. Unlike hard computing, which requires an accurately stated logical, systematic and consistent model or a complicated computational environment, a soft computing technique derives its power of generalization from previous outputs from previously learned inputs. They have shifted the aim of computing from classical techniques to the human mind, which has a remarkable ability to store and process data that are extensively vague, uncertain and lacking in category. Soft computing employs its methods in a complimentary rather than a competitive way. One important combination is a genetic fuzzy system (GFS), which combines the benefits of fuzzy logic and the genetic algorithm. Fuzzy logic is used to form the basis of representing different forms of system knowledge or model the interaction and relationships among the system variables where classical tools are unsuccessful. Genetic algorithms are heuristic search techniques known for their strong search capabilities in multifaceted search spaces. The algorithm presented here is an optimized protocol designed using a GFS system. The genetic fuzzy system is used to select optimal nodes that can become CH and also an optimal number of CHs so that complete network coverage is achieved whenever clustering is initiated. In OZEEP once the task of clustering or cluster setup with associated cluster

heads and cluster members is over then the steady state phase starts with data aggregation from cluster members and data routing between CHs towards sink or base station. The routing algorithm used is Zone Based Energy Efficient Routing Protocol (ZEEP). After the data transmission is completed re-clustering is initiated where GFS is used to select the best nodes based on the above mentioned network parameters and current network status to be CH for the next round. LEACH, HEED, APTEEN, LEACH-C, PEGASIS [1] are all dynamic clustering algorithms which have a cluster setup phase where a CH is selected and a steady state phase where the selected CH is maintained and data transmission is performed. We briefly present in Table 1 and 2 some works done on clustering in the literature using fuzzy logic and the genetic algorithm. Based on our understanding, we outline some comparative points of the fuzzy and genetic algorithms before we use a combination of the two techniques to complement each other. We then present some past studies on GFS to show the benefits of using a blend of the two techniques to give a GFS-based methodology. We briefly summarize the characteristics and limitations of each technique in Table 3 so that we can draw attention to domains where each technique compliments the other.

Page 7 of 36

Table 1 Literature survey on the fuzzy system used for clustering in wireless sensor networks Parameters considered

Anno et al. (2007) [8]

Fuzzy Logic (FL)

Distance, energy and network traffic

Taheri et al. (2012) [7]

Fuzzy Logic (FL)

Clustering, Cluster Head(CH) selection for increased network lifetime On demand Clustering, Cluster Head(CH) selection for increased network lifetime

Godbole (2012) [26]

Fuzzy logic(FL)

Distance and residual energy

Bagci et al. (2013) [5]

Fuzzy Logic (FL)

Song et al. (2013) [15]

Fuzzy logic(FL) and Ant colony optimization(A CO)

Clustering, Cluster Head(CH) selection for increased network lifetime Unequal clustering, Cluster Head(CH) selection for increased network lifetime Cluster Head (CH) selection by fuzzy logic and optimal inter-cluster done by ACO for increased network lifetime

Khoshkangi ni et al. (2013)[20]

Fuzzy Logic (FL)

Hashemi et

Fuzzy Logic

Output expected

Issues not addressed

Cluster head

Mobility and node density not considered

Fuzzy cost, which is a value based on the fuzzy if-then rules. A smaller cost means a high chance of becoming a cluster head Radius estimation to select final CHs

Mobility and complete coverage not considered

M

Triangular

Mobility, node density and degree of node not considered Mobility, node density and degree of node not studied

Distance and residual energy

Triangular

Radius estimation to select final CHs

Distance, residual energy and node density

Triangular

Mobility not considered

Residual energy and centrality

Gaussian

FL outputcandidates to be cluster head and final selection output based on estimation of CH competition radius ACO output optimal routing path Cluster head based on high levels of energy and good centrality`

Distance,

Gaussian

Best nodes to

Mobility

ed

pt

Ac ce

Clustering, Cluster Head(CH) selection for increased network lifetime Clustering,

Triangular

an

Remaining energy, node degree and node centrality

Membersh ip function (MF) used Trapezoid al and triangular

ip t

Aim

cr

Soft computing technique used

us

Author and year of publication

Mobility, node density not considered

Page 8 of 36

Type-2 Fuzzy logic and Ant colony optimization(A CO)

Cluster Head selection by fuzzy logic and optimal inter-cluster done by ACO for increased network lifetime

Triangular

1. First step fuzzy output chance or caliber of a node to be a gateway among other competing nodes 2. second step fuzzy outputchance of nominated gateways to be elected as CH FL outputcandidates to be cluster head and final selection output based on estimation of CH competition radius ACO output optimal routing path

Mobility not considered

ip t

Zhang et al. (2014) [18]

Fuzzy is used in two steps: 1. Gateway election based on distance, energy and centrality 2. CH election based on concentratio n, distance and efficiency Distance, energy and node density

not considered

cr

Fuzzy Logic (FL)

be cluster head

us

Arya et al. (2014) [19]

energy, traffic and node density

Primary MFs Trapezoid al and triangular secondary MF – interval type-2

an

Cluster Head(CH) selection for increased network lifetime Clustering, gateway cluster head(CH) selection for increased network lifetime

M

(FL)

Mobility not considered

pt

ed

al. (2013)[22]

Ac ce

Table 2 Literature survey on use of genetic algorithm for clustering in wireless sensor networks Author and year of publicatio n Wazed et al. (2007) [25]

Soft computin g technique used Genetic Algorith m

Issues addressed

Parameters considered

Disadvantages/ Issues not addressed

Routing, scheduling data gathering at cluster heads for increased network lifetime

1. Fitness function:

Mobility not considered

Lnet= Einitial Emax where Lnet is network lifetime based on rounds Einitial is the CH initial energy Emax is the maximum energy used by any CH 2. Selection-roulette wheel 3. Crossover- uniform crossover(swapping probability 0.5) or k-point crossover selected randomly

Page 9 of 36

Vinayak et al. (2010) [17]

Simulate d Annealin g- local search efficienc y and Genetic Algorith m – for rapid global search

Ac ce

Miao et al. (2009)[27 ]

Genetic Algorith m

cr

ip t

Mobility not considered

us

1. fitness function -

an

Genetic Algorith m

2. Selection- ranking selection 3. Crossover- sector by sector or singlepoint crossover 4. mutation (pm= 0.1)

M

Seo et al. (2009)[6]

Selection of a cluster head that can minimize the maximum intra-cluster distance between itself and the cluster member and the optimization of energy management of the network Design of a twodimensional genetic algorithm for clustering that returns the location information of the nodes Energy efficient routing algorithm based on selection of high energy nodes as cluster heads

Fitness function based on energy and node position

ed

Simulate d Annealin g and Genetic Algorith m

Only transmission distance between nodes is considered. No actual energy consumption or mobility studied

Mobility not considered

pt

Jhang et al. (2008) [10],

4. Mutation- node in the chromosome, which dissipates maximum energy denoted as critical node 1. Fitness function f (x) = 1/ cost 2. Selection-roulette wheel 3. Crossover- uniform crossover (pc =0.85) 4. Annealing mutation (pm= 0.1)

To study the random behavior of nodes which are either a cluster member or a cluster head. Their energy usage in data communicati on is studied based on

2. Selection- Elitist selection 3. Crossover- two point crossover (pc=0.85) 4. mutation – 0.1 1. Fitness function

Actual clustering not discussed. Mobility not considered

2. Selection-not mentioned 3. Crossover- single point crossover 4. mutation - 8th bit of the chromosome is flipped

Page 10 of 36

us

cr

2. No description on chromosome representing individuals is given. No matter is given on the genetic operators’ selection, crossover or mutation.

ip t

Mobility not considered

an

Genetic Algorith m

1.Fitness Function

1.Fitness function-

Mobility not considered

ed

Rahmani an et al. (2011) [2]

Genetic Algorith m

M

Babaei et al. (2010) [21],

based on varying the message size and the distance between the nodes. To select optimal nodes as CHs and optimal number of CHs for uniform distribution of nodes in the network. The clustering parameters selected are distance, network lifetime, number of cluster members. Clustering; genetic algorithm is applied to LEACH-C protocol to determine optimal node and optimal number of cluster heads.

Ac ce

pt

2. Selection - tournament selection 3. Crossover- single point crossover (pc = 0.8) 4. mutation (pm= 0.1) 1. Fitness function- based on communication energy, CH fraction and the CH speed

Sarangi et al. (2011) [24]

Genetic Algorith m

GA-based clustering with mobility, cluster head selection based on low mobility

OptimScore = (ME x WME+ CHFrac x WCHF + CHSpeed x WCHS) 2. Selection-roulette wheel 3. Crossover- single point crossover(pc =0.8) 4. Mutation- 0.3 5. Mobility model - random waypoint

Periodic in nature and runs on gateway nodes, location to be known a priori, nodes that move away from a cluster become a CH and unfortunately perform direct transmission to the BS until reclustering happens

Page 11 of 36

Propose two algorithms that improve network energy consumption by partitioning the network into clusters.

1. The CHs are selected based on the position information and the residual energy of the node. The fitness function will aim at minimizing the selection of a node with low residual energy like a CH. The number of CHs is constant in each round, giving a set of stable clusters. 2. No explicit mention of the fitness function is provided. The authors compare P-LEACH, HS-LEACH, and GP-LEACH in their paper. 2. No details are provided in their work about the operators’ choice

Table 3 Table of comparison between Fuzzy Logic and Genetic Algorithm Disadvantages

1. Can handle problems with incomplete information and arrive at real time decisions

1. The method is logic based and may create a degree of inexactness

2. Needs prior knowledge and understanding of the problem in generating the knowledge base

M

2. Considers the ‘gray area’ and believes that not everything is true or false

ed

3. Provides  Flexibility  Options  Allows for observation

4. Can provide piece-wise linear approximation to a system

pt

5. Stores knowledge explicitly

Ac ce

6. No learning required

7. Gives same importance to all of the parameters

When to be used

us

Advantages

an

Soft Computing Technique Fuzzy Logic

Mobility not addressed

ip t

Genetic Algorith m and Harmony Search

cr

Karimi et al. (2012) [12]

1. Situations where mathematical solutions cannot be easily defined 2. Situations where information representation and processing is based on human-oriented knowledge

3. Non-linear systems cannot be approximated. 4. Can be slow 5. Does not perform optimization

Page 12 of 36

2. Based on natural selection and biological evolution 3. Once the model is fixed, irrespective of whether the model is rule based, equation based or a class of function, it converges within the search space even if provided with lesser data. 4. Does not store knowledge as it generates optimal solution with the best survivors or values of the previous solution

1. Highly dependent on the model or objective function selected. 2. A poor model can lead to a poor result with optimization holding no meaning 3. Problem may have multiple parametric solutions

5. Has ability to learn

3. Reasonable computation requirement 4. Multifaceted design of engineering or complex search space

us

5. Mutation interference

Ac ce

pt

ed

M

7. Can handle applications with non-linearity

4. Local minima and premature convergence

2. Dynamic process control

an

6. Allows multiple parameters to form the optimization objective function with different levels of importance given to each parameter if required.

1. High rate of convergence

ip t

1. Derivative free optimization

cr

Genetic Algorithm

Page 13 of 36

an

us

cr

ip t

Abedini et al. [23] have proposed a two-level clustering algorithm. In their work, the fuzzy module is used to select CH candidates based on energy, density and centrality of the node. In the second phase, a GA is used to make the final selection of CH. However, their work does not consider node mobility. Looking through the presented literature, it is obvious that given a coverage area and variable mobility levels between nodes, choosing a cluster head is challenging. If two CHs are closely situated, the chances of clusters overlapping with each other are high, and in such a scenario, the nodes may belong to multiple clusters. This drains unnecessary node energy and also impacts the network lifetime. Similarly, in some cases where nodes are sparsely located, they may not be part of any clusters. In such situations, a direct communication between these nodes that act as a CH and the base station occurs. The nodes consume more energy in reporting their sensed data as the distances involved increase. This greatly impacts the node’s lifetime and as the energy of the node decreases, the chances of network partitioning also increases. As the intricacies of CH selection in such scenarios increase, it becomes essential to design techniques that concentrate on finding the optimal number of capable cluster heads that can be well distributed in the network to achieve maximum coverage and maximum network lifetime. We now proceed to introduce the proposed GFSbased model for clustering in ZEEP. We have implemented the proposed algorithm on our existing routing protocol ZEEP designed for mobile sensor networks and have tried to optimize its CH selection and clustering process. Our previous design for ZEEP [9] supports a two-tier design for data aggregation and dissemination in sensor networks. It uses the mobility factor of a node as a criterion to be selected as a CH. However, the algorithm ZEEP does not provide a solution for nodes that have the same mobility factor in a particular cluster. In our proposed algorithm, the CH selection process is optimized to allow nodes that have similar inputs for energy, distance, node density or mobility levels to be evenly distributed throughout the network so that the overall energy consumption of the network is minimized. The algorithm works in two phases. In the first phase, all nodes use the fuzzy inference module, which is inbuilt in them. This module helps each node to decide on their caliber to contend for the CH role. The decision is based on their residual energy, number of nodes around them or node density, their distance from the base station and finally their speed or mobility factor. The output of the fuzzy system, which is the first level of CH nomination, is sent to the base station that runs a GA. Only the nodes that are nominated to contend for the CH election are represented in the chromosome. This helps in avoiding unwanted participation of all of the nodes

Ac ce

pt

ed

M

Reading through the contributions of each technique in problem solving, we can say that the fuzzy system is capable of solving a variety of multifaceted issues. However, their limitations in handling certain issues have led researchers to adopt different approaches to enhance the functioning of the fuzzy system. One such idea is to hybridize the fuzzy model with either genetic algorithms or neural networks to obtain a genetic fuzzy or a neuro-fuzzy system. In our work, we deal explicitly with a genetic fuzzy system that is an augmentation of GA with Fuzzy. The ability of GA to learn and solve non-linear and derivative-free optimization problems and fuzzy logic’s capability of identifying and establishing relationships between system variables can provide excellent solutions to real world applications where other search techniques or tools are unsuccessful. A look into the literature and past research can give us an overview of how to use GFS in communications and networking. Filho et al. [4], propose a fuzzy clustering-based fitness evaluation model to improve GA performance in terms of execution speed, complexity and quality of solution. Each similar individual is grouped into clusters using fuzzy c-means and participating learning. The algorithm is designed to reduce direct evaluation of fitness functions to speed up computation. Leal et al. [11], proposed a route selection algorithm for WSNs using GFS. The proposed algorithm is based on the concept of many-to-many communication. The algorithm selects the most appropriate sink among many possible sinks. This selection is performed with the help of a fuzzy decision tool that gets energy and number of hops as inputs. Based on the favorable outputs, GA selects the most appropriate sink and route to reach this sink. GA is used to tune the Fuzzy Inference System so that it generates optimal outputs. Esmaeil et al. [3] have proposed a clustering scheme based on fuzzy logic and GA. The clustering decision is taken by the base station, which runs a genetic algorithm to select the cluster heads. The BS receives information about the CH candidates from each node that has a fuzzy module embedded in it. Once clustering is completed, the BS broadcasts this information to the network, which helps all nodes to decide on their clusters and corresponding CHs. However, node mobility has not been considered in their design. Omari et al. [14] have proposed a two-phase clustering technique. In the first phase, the authors use the fuzzy logic system to separate the network into smaller domains and then perform clustering. For clustering, authors run GA on base stations that are assumed to have surplus computational and energy resources. Once the clustering is done, the authors use a token ring in their algorithm, which runs in each cluster to rotate the CHs based on residual energy.

Page 14 of 36

and also helps the algorithm to converge faster to the finest solution. The success of using a combination of the genetic and fuzzy systems results in a balanced and uniform selection of nodes such that the energy consumption of the network is reduced, leading to an increased network lifetime. We now formally define the proposed algorithm.

iii.

Nodes with a lower mobility factor stand a better chance of becoming a CH than nodes that are highly mobile.

iv.

If two nodes have the same mobility factor, the node that has more one-hop neighbors becomes the CH candidate.

3.System and Energy model 3.2. Energy Model

A gateway node or a SINK node is assumed to be a hub of a powerful processor and unlimited energy that is responsible for classifying the network into a set of balanced clusters.

iii.

Our design assumes that every node has the same energy and capability.

iv.

The algorithm assumes that the location of all nodes is known to the SINK/gateway node.

ip t

ii.

cr

The algorithm can start on a network with randomly distributed nodes or with a random set of cluster configurations in the beginning.

ed

M

i.

We know that a radio transceiver basically has two tasks to execute: transmission and reception of data between nodes. Since sensor nodes are energy constraints so it becomes extremely necessary to provide low energy consumption mode to transceivers which can allow them to be turned off for most of the time and be activated whenever data has to be received or transmitted. When we study the amount of energy consumed during transmission we see that in principle the energy consumption is because of two sources: one due to the RF signal generation that depends on the distance and the modulation technique used also known as transmission power emitted from the antenna and second due to the energy required by the electronic circuitry for frequency synthesis, conversion, filter. The transmitted power is dependent on the system characteristics like BER, distance, multipath or free space model. It is generated by the amplifier of the transmitter and therefore amplifier’s power consumption, depends on the power it is supposed to generate for the required signal strength. The energy model used in the proposed work for OZEEP considers the energy dissipation of a node in both transmissions as well as the reception state. The energy parameter includes energy usage to run the transceiver circuitry, Eelec, energy used by the transmitter amplifier, Efs, to send a message of l bits to a distance d2 (distance ≤reference distance d0 (free space model)) [3] [24]. Therefore, the energy model for transmission of data is given by

us

The proposed system under consideration has the following characteristics:

an

3.1. System Model

For our model, we have fixed the transmission ranges for cluster members (CMs) and CHs.

ii.

iii.

pt

i.

Ac ce

3.1.1. System Model for stationary network

Nodes can function at two power levels; high or low or a blend of high/low. The maximum transmission range defined for a CM is CMRmax and for CH is given by CHRmax [24].

3.1.2. System Model for mobile network i.

ii.

Nodes can regulate their transmission power according to their distance from the proposed CH by CMRmax and the distance of the CH to the gateway or the sink by CHRmax. As soon as a node discovers that it has moved out of range of its CH by a distance of more than R, it behaves like a CH until re-clustering happens again.

ETx = lEelec+lEfs d2

(d≤ d0) (1)

And the energy usage model for receiving l bits is given by ERx= lEelec

(2)

where only the electronic circuitry of the transceiver is on to receive the incoming bits. 4. Proposed Work The proposed algorithm is a two-tier model with cluster heads at the first level and cluster members in the second. The CHs are responsible for data aggregation and transmission of aggregated data to

Page 15 of 36

Method

depict

Final selection of CH from the lis t provided by Fuzzy logic output

ed

construct

construct

Ac ce

Standard deviation

Fuzzy simulation in MATLAB

ip t

Distance, node density, energy and mobility of node (o/p: optimal number of nodes to be selected to contend for CH

Less/more number of modes selected to be CH, imbalance energy usage of the network reflected

depict

Best node selected as candidate to contend for CH role

depict

Final selection of nodes for CH role completed and an optimized network generated

Level 4

Optimized network fed into network simulator ns2 and studied against state of the art routing algorithm for WSN

depict

Mathematical model

Model solvi ng

Level 3

Genetic Algorithm simulation in MATLAB

Intuitive model

Distance, node density, energy and mobility

pt

Level 2

Expected: Optimal clustering with optimal or best nodes selected to be CH

cr

depict

Level 1 Genetic Algorithm

us

Result

M

Fuzzy variable

then declare themselves as candidates for CH to the base station. The base station, which runs a genetic algorithm, converts this information received from the candidates into chromosomes and makes a final selection of nodes to take up the CH role. Although the task of the genetic system is reduced as a result of using a fuzzy system, the role of the genetic algorithm is still vital as its task is to carefully classify the nodes for the CH role from the pool of available candidates and organize these critical nodes to form balanced clusters in such a way that the network consumes a minimum amount of energy. Another benefit of adopting this technique for CH selection is that it allows only CH contenders to communicate with the base station, thus reducing on the number of nodes reporting to the base station and therefore avoiding collision among nodes. Figure 8 gives the problem formulation and the scheme for solving the CH selection and clustering process carried out in this work.

an

the base station via multiple hops. However, this communication is restricted to cluster heads and is known as inter-cluster communication. The second level or intra-cluster communication, as it is prominently known, is a single-hop communication because each cluster member reports its sensed data to the corresponding CH of the cluster these members belong to. The presented algorithm performs clustering in two steps. The first step is a screening process that identifies capable nodes among all of the deployed nodes to contend for the CH role. This is done with the help of the fuzzy module built in the nodes. Formerly, when the algorithm was initialized, each node entered its energy, distance, node density and mobility details as inputs to the built-in fuzzy inference module. The fuzzy inference engine is responsible for mapping the entered inputs to generate an output that shows whether the node can be chosen as a CH. The best nodes based on the output of the module are nominated as a cluster head candidate by use of a timer [3], which starts based on the result of the fuzzy system and reaches a value zero, faster for nodes with better utility. These nodes

Optimal energy consumption of the network due to optimized clustering, minimal packet loss achieved

Model solvi ng

Fig.8 Problem formulation and model solving in the proposed work Mobility is an important criterion selected in our proposal as almost all real-time applications these days employ mobile nodes with variable speeds.

These nodes can move around the application domain and carry out vital sensing, reporting and processing. It becomes necessary for researchers to study their

Page 16 of 36

an

us

cr

ip t

movement. Once a node has a defined cluster and a CH then any movement out of the selected CH range will lead to alienation and the node itself starts a direct transmission with CHRmax towards the base station until re-clustering happens. We now implement this clustering methodology on the existing routing protocol ZEEP, where we aim for an optimized clustering and CH selection to reduce the network energy. The optimized ZEEP, which we obtain after introducing the GFS system, shows an optimal number of well-distributed CHs. The optimized network of OZEEP, as we call it, shows good results in terms of network energy and the packet drop ratio when compared with ZEEP. Further, our study in this paper also incorporates a study of available membership functions (MF) suitable for this environment before we made a choice of triangular MF. The work also includes a study on Mamdani models and Sugeno models to decide on the FIS. We also structured and studied GA to obtain CHs in the network without incorporating a fuzzy system into it. Each technique was computed separately before presenting the final outcome of the combination of the two methods to the world. The idea is to observe the behavior and results of each technique and to design systems with combinations of the two methods. Before we describe our study in detail, we would first like to present the process flow chart representing the details of implementation of our proposed model in figure 9. This would help the readers to understand the selection process easily.

Ac ce

pt

ed

M

behavior and hence suitably allocate roles to them during clustering. In this context, it is observed that nodes that have high mobility have less chance of becoming a CH as they would not be able to form steady clusters. As a result, our presented algorithm looks for nodes that have low mobility levels before it elects them to be a CH. Nodes with low speed as CH form stable clusters such that the standards selected for network parameters do not change noticeably during the clustering process and energy efficient data routing is accomplished without significant disturbances. Conversely, the algorithm does not place any restrictions on the speed or movement of non-cluster heads. This means that these member nodes can move with any speed up to Vmax in any direction. However, there is a possibility that some nodes that have high mobility when compared to a predefined threshold value (distance more than R) might move out of their clusters and stroll in isolation. In such scenarios, these nodes adopt a direct transmission with a CHRmax range to the base station until re-clustering happens. It is possible sometimes that the mobile node may not move away very far and might receive signals from a nearby cluster. In such cases also we have assumed that the registration of this new node does not happen with the new cluster and cluster head until re-clustering happens. This is to ensure that the steady state phase (data routing) is executed without any disturbance. It is important to note in which phase the cluster member moves away. At the time of clustering a node registers with a CH which requires minimum communication energy. However, if the node moves away at the beginning just before choosing the current CH, then it is required to choose a CH which is closest to it depending on its speed and direction of

Page 17 of 36

Start

Fuzzy module embedded in each deployed sensor node Local information gathering from nodes (distance, node density, residual energy & mobility) Crisp input

Fuzzification

Fuzzy rule base (ifthen rules: AND operator: Min, OR operator: Max

Output of FIS sent to BS which runs GA

Crisp output

Fuzzy Output

1. Fitness function defined 2. Chromosome representation

Nodes selected as cluster head candidates

Defuzzification (Centroid method)

Initialization (Uniform random choice)

Population

cr

Fuzzy decision block FIS (Mamdani Inference System)

ip t

Fuzzy input

Survivors selection

Offsprings

us

Membership function (Triangular MF)

Yes

an

Mutation = 0.3

Termination condition satisfied ? (Cluster construction-optimized network with optimal CHs obtained)

Crossover (Single point crossover with crossover rate 0.8) No

M

Stop

Comparison of GFS optimized network with ZEEP in terms of packet drop and energy

GFS based optimized network shows improved results

Optimized network after clustering fed into network simulator ns2

Parents Selection (roulette wheel)

ed

Fig.9 Process flow and implementation details of the proposed algorithm OZEEP 5. Design of fuzzy inference engine module

Ac ce

pt

We propose the Mamdani inference engine for our system. The module consists of a Fuzzy Logic Controller (FLC) [3] that determines the best nodes as contenders for the CH. This controller is made up of components such as the fuzzy inference system, fuzzy decision rules, fuzzifier engine and defuzzifier. The FLC takes four input parameters, namely, i.

Energy of each node (the higher the energy, the better is the chance of becoming a CH)

ii.

Distance of a node from the base station (the lesser the distance between the node and the BS, the better is the chance of being a CH).

iii.

iv.

Node density (to see how many neighbors a node has so that its influence on these neighbors can be checked to decide on it being a CH) Mobility level of each node (the lower the speed or velocity of movement of the node, the higher the chance to be a CH).

The linguistic variables selected for each of the inputs are: i.

Energy = {vlow, low, med, high, vhigh}

ii.

Distance = {near, med, far}

iii.

Node Density = {low, med, high}

iv.

Mobility = {stop, slow, med, fast}

The output of the membership function for the linguistic variable would be: OUTPUT = {vsmall, small, med, long, vlong} The fuzzy operators selected for the FIS are: i.

And operator: ‘Min’

ii.

Or operator: ‘Max’

iii.

Defuzzification operator: centroid method

The fuzzy rules applied on each node would output the following decisions:

Page 18 of 36

If (energy is vhigh) and (distance is near) and (density is high) and (mobility is slow) then (output is vsmall)

iii.

If (energy is high)and (distance is near) and (density is med) and (mobility is med) then (output is small)

iv.

If (energy is med)and (distance is med) and (density is med) and (mobility is stop) then (output is vsmall)

Figure 10 shows the FIS module selected for our proposed algorithm OZEEP.

The optimized ZEEP consists of a fitness function and parameters that optimize ZEEP operation in the network. The algorithm OZEEP aims at complete network coverage by generating stable and balance clusters in the network. GFS which is a two-step process uses genetic algorithm to generate the best nodes as cluster heads and also an optimal number of such CHs so that these nodes could be distributed optimally in the network for complete coverage and balanced cluster formation could be realized. The input to the GA module is the resultant of the first level screening done by the fuzzy module; list of nodes that can contend for the CH election. On this input genetic algorithm parameters are run to produce a desired solution.

ip t

ii.

6. Genetic algorithm parameters

cr

If (energy is low)and (distance is near) and (density is low) and (mobility is fast) then (output is vlong)

us

i.

6.1. Chromosome Representation

M

an

Once the base station decides on the CHs, it broadcasts the result in the entire network. The entire network is represented with a single chromosome of length n, where n is the number of nodes in the network. A ‘1’ in the chromosome means a CH, whereas a ‘0’ represents a node that will select a CH according to CMRmax to be part of that cluster.

ed

6.2. Genetic Algorithm parameters

Fig.10 Fuzzy Inference System for OZEEP

Ac ce

pt

The fuzzy decision rules help nodes decide whether they stand a chance to become a CH. For example, a node that has very high (vhigh) energy, has slow movement, is near to the base station and influences or has a larger number of neighbors stands a very high chance of being selected by the base station as the CH. We have used a Mamdani model for decision making in our algorithm as against the Sugeno model. Both techniques are almost similar in their functioning; however, there still exist few differences that make it possible for the designers to make a choice of the inference model depending on their application domain. Sugeno models use a weighted average method to compute the output, which is in a crisp form. The consequent rules are not fuzzy here, and the outputs generated are either constant or linear. Mamdani models, however, allow expertise to be described in a more intuitive, human-like manner. They are more expressive and interpretable. These models use a defuzzification technique to generate an output, which is also in the form of a membership function. Figure 10 shows this relation. They are well-suited for decision support applications.

The proposed algorithm uses a single point crossover so that the interchange of segments results in a new cluster configuration that has optimal outcomes in terms of energy. Mutation is a process of randomly changing one or more bits of a chromosome to generate a new chromosome. The mutation rate selected should be kept low so that node roles do not change much from being a CH to a CM. The selection process helps choose chromosomes that have the capability of giving better next generation chromosomes. The technique selected for our algorithm is roulette wheel selection so that the fitness value of each chromosome is normalized to 1, which is interpreted as a chromosome standing a higher chance of being selected for next generations. 6.3. Fitness Function GA is always equipped with a fitness function, which enables it to score and rank individuals. The fitness function is task specific, and it evaluates the caliber of a chromosome as a solution for further generations. It is believed that a higher value of fitness represents a better individual for creating superior offspring in the search space. A fitness function defines the objective of the problem and, hence, is suitable for accommodating problems with a single objective or multiple objectives. However,

Page 19 of 36

6.3.1 Cluster head fraction (CHFrac)

us

cr

ip t

Hierarchical topological networks can achieve a good energy efficient control by deliberately limiting the number of active nodes and links in the network. This can be done by turning some nodes off and using only crucial links while ignoring the other ones. Clusters are such subsets of nodes that can reduce the number of active nodes to (VT ⊆ V) and active links to (ET ⊆ E), where V is the set of all nodes in the

Ac ce

pt

ed

M

With design, study and analysis of the proposed algorithm it is evident that mobility dominates other design parameters. Therefore, it is required that the speed or mobility level of CHs be taken as a parameter to compute fitness function. Another factor is the optimum number of CHs. Percentage of CH is important as both large number of CH and a smaller number of CH would drain the network of its energy. This can be understood from the total number of nodes in the network that could be fractioned into smaller subsets where their group leader was just one hop away. Hence, it was necessary to study the required CH to CM ratio for creating balanced clusters. The variable CHFrac in the objective function helps find this desirable CH to CM ratio that can give stable cluster with complete network coverage and minimum energy consumption of the network. Finally it is necessary to curb the energy consumption of the network due to data transmissions. Variations in distance from the base station and cluster head lead to distinct power levels to be adopted for data transmission. This leads to high energy drainage of nodes and hence, smaller network lifetime. Thus it is important to calculate the required communication energy in the system and to keep this value as minimum as possible. It was important for genetic algorithm to consider all the three variables, namely, mean communication energy, CH fraction and CH speed to be studied as objective function. Genetic algorithm tries to minimize all the three parameters so that the best CH candidates nominated by fuzzy module are selected for final CH role. In the proposed algorithm, we select a variable fitscore, which is computed by the fitness function, whose objective is to select a chromosome that minimizes the fitscore value. In other words, the lower the fitscore value, the higher is the fitness of that individual.

The function fitscore defined in the proposal comprises three parameters or objectives that we try to minimize. The fitscore is proportional to the constraints and is calculated by summing them. These parameters are Cluster head fraction (CHFrac), Mean Communication Energy (ME), and Total Cluster head Speed(CHSpeed) [24]. We define each of these variables below:

an

most of the real world problems are made up of multiple conflicting objectives, and therefore, optimizing the fitness value with respect to one objective may result in unacceptable solutions with regard to other objectives. Hence, it is difficult to design a multi-objective fitness function that optimizes all of the objectives concurrently. A realistic approach is to look through a set of solutions, where each of the objectives is satisfied to an acceptable level without any domination from other participants. This technique of moving from one acceptable solution to another with tradeoffs between conflicting objectives is known as the Pareto optimal set of solutions. They are unlike the weighted sum method where the accurate selection of weights is a challenge as minute perturbations in the choice can result in entirely different solutions. Pareto optimal solutions are more practical for real world problems, and therefore, the solution of the multi-objective fitness function defined in the proposed work is based on a Pareto optimal solution set.

network and there exists an edge (v1, v2) ϵ E ⊆ V2 if v1 and v2 can directly communicate with each other. This means that clustering converts the network a graph T = (VT, ET) such that VT ⊆ V and ET ⊆ E. Or

in other words, the idea is to work out a modified graph T from a graph G representing the original network G. However, even after dividing the network into smaller subsets, these clusters collectively incorporate all nodes of the graph. A crucial problem in clusters is to compute cluster heads; a representative of clusters. A critical parameter is percentage (%) of CH. Energy dissipation of the network varies as the percent of nodes that are CH is changed. If all nodes are selected as CH that is, at 100% CH, then it is same as direct transmission. If the number of CH is low then the anticipated distance between cluster members and CH is more, and hence, the members will have to expend more energy to transmit to CH to maintain the expected Bit Error Ratio (BER). However, if CH number is high then the consequence is a more expensive transmission from CH to the BS. If the average number of nodes in a cluster has to be minimized so that every member is just one hop away from its CH then it means we need to find a maximum (dominating) independent set which

Page 20 of 36

N = |CH| + |CM|

(3)

The Total Cluster Head Message Energy (TCHE) denotes the amount of energy used to transfer one message from a cluster head to the sink. It is represented as .

CHFrac = |CH|

Ac ce

N

Mathematically we represent these as TCME

=

pt

ed

M

If we separate the elements of N in two groups we get a subset CH which is the set of all cluster heads and CM is the set of all cluster members in the network. For a desirable cluster configuration it is important that a good CH to CM ratio be preserved. We now describe a CHFrac that defines an optimal number of CHs required in the network. It is given by the equation

ip t

E. This maximum independent set contains all probable nodes without defying the maximum independence property and therefore result in several clusters. They also naturally partition the network with a subset of nodes that can control the whole network. Once this set of CHs is created it is important to find a desirable cluster configuration which achieves complete network coverage using minimum network energy and resulting in an increased network lifetime. For a formal definition we consider a network with N nodes,

We now advance to the next definition, which is the Mean communication Energy (ME). This component models the total energy usage of the system for data delivery and is a trivial component in representing the main energy consumption of the system. It is represented by two strictures, namely, the Total Cluster Member Message Energy (TCME) and Total Cluster Head Message Energy (TCHE). The TCME parameter is further portrayed in two parts. The first part describes the total amount of energy used to send one message by each CM to its corresponding CH. In other words, it involves the energy spent during transmission and reception of a message by a CH and a CM and is called as Communication energyi,j. The second element TCHE represents the total energy used by a CH in combining all messages received by its corresponding CMs into its own message for the sink. This is the data aggregation energy (Aggregation energyj).

cr

are joined by an edge in E – ∀ c1, c2 ϵ C : (c1, c2) ∉

6.3.2 Mean Communication Energy (ME)

us

C : Ǝ c ϵ C : (v, c) ϵ Ε and further no two nodes in C

an

contains a subset C, where C ⊆ V such that ∀ v ϵ V –

desirable cluster configuration that could partition the network in a way that all nodes are covered at the same time network’s energy requirement is minimized.

(4)

where, CHFrac is obtained by dividing the total number of elements or the cardinality of set CH by total number of nodes N, in the network. As an example, if N=100 and cardinality of set containing all cluster heads CH = 20 then CHFrac = 2 : 5, that means for two cluster heads there are 5 nodes as cluster members. In other words it is possible these clusters have common nodes and hence the two clusters overlap. In a dense network, we sometimes cannot avoid overlaps but the requirement is to find the right number of CHs so that we do not create too many CHs or too less CHs and impact the network’s energy consumption. It is possible that the fuzzy system might result in a large set of nodes for CH position. However, the variable CHFrac tries to estimate an optimal ratio of a CH and the number of one hop away cluster members it can support for a

(5)

TCHE =

(6)

Once TCME and TCHE are defined, we can now give the principle for Mean Communication Energy, which is TCME ± TCHE ME =

N

(7)

However, for a good energy efficient network configuration it is important we try to understand Mean communication Energy (ME) and its components in more detail. Let us assume an area of dimension A with N sensor nodes deployed randomly, and distance of nodes defined from base station or cluster head with reference to distance d0 (whether multipath (d>d0) or free space model (d≤

Page 21 of 36

ECHRX = l ERX CM

(8)

The energy required by cluster head to aggregate the message sensed by CH itself and the same message received by it from its cluster member is given as Aggregation energyj = lEAE (CM +1) (9) Therefore, the total energy consumed by a CH in a round which includes the energy spent to receive, aggregate and send a packet to the base station is given by ECH, where

6.3.2.2 Modeling multi hop transmission In multiple hop data transmission technique, we assume that CHi transmits data to CHj and then in turn CHj transmits the received data from CHi to the base station. ECHi to CHj = l ERX CMi + lEAE (CMi +1) + lETx + lEfs d2CHi to CHj (16) ECHi and CHj to BS = l ERX (CMi + CMj) + lEAE (CMi + CMj + 2) + lETx + lEfs (d2CHi to CHj + d2CHi to BS) + l ERX (17)

ip t

d0)), then, the energy used by a cluster head to receive l length of message from its cluster members (CM) is given by-

6.3.2.1. Modeling direct transmission

6.3.3 Total Cluster head Speed (CHspeed )

us

(10)

We can write equation (10) as ECH =lERX CM + lEAE (CM+1)+lETx + lEfs d2CH to BS

ed

M

(d≤ d0) (11)

an

ECH = ERX + EAE + ETX

cr

where, lETx is the power required by the CH to transmit l length of data packet, Efs is the energy consumed by the transmitter amplifier (free space model) and dCH to BS is the distance between the CH and the base station in free space model.

Equation (17) shows that the energy consumption in multi hop communication is less when compared to direct transmission. The reasons being the distance metric which is different for inter cluster communications. The distance parameter is an important factor that requires the transmission power of the nodes to be regulated. Distance variation becomes inevitable when nodes are mobile. Therefore, our algorithm OZEEP tries to create stable clusters by identifying slow moving nodes to become CH. The requirement is that the cluster disruptions due to mobility are less and therefore the Mean Communication Energy which impacts the network energy consumption severely can be kept to the minimum.

pt

We assume two clusters with cluster heads CHi and CHj and corresponding cluster members CMi and CMj. If data is transmitted directly to the base station by these CHs then the total energy consumed by the network in free space model is given as

Ac ce

ECHi = lERX CMi+ lEAE (CMi +1) + lETx + lEfsd2CHi to BS (12) ECHj = l ERX CMj + lEAE (CMj +1) + lETx + lEfs d2CHj to BS (13) ECHi + ECHj = l ERX (CMi + CMj) + lEAE (CMi + CMj + 2) + 2 lETx + lEfs (d2CHi to BS + d2CHj to BS ) (14) If the two CHs; CHi and CHj are closed to each other, then their distance to the base station would be approximately same. In that case the total energy consumption of equation (14) can be written as ECHi + ECHj = l ERX (CMi + CMj) + lEAE (CMi + CMj + 2) + 2 lETx + 2lEfs (d2CHi to BS) (15) Equation (14) shows that energy consumption for direct transmission is high and is twice the amount of energy spent by a single CH.

Cluster heads are representatives of clusters and are therefore responsible for multiple activities like data aggregation from cluster members, data routing, and maintaining network topology; consequently, energy dissipation by a CH is more when compared to a cluster member. It, therefore, becomes necessary to rotate the role of a CH so that there is balanced energy dissipation in the network. In other words, any node which has better capabilities depending on the criteria set by the network can become the CH in subsequent rounds. If decision criteria are distance, mobility, node density and residual energy of the node then it is desirable that a CH nominee should have least distance from the base station, slow moving node with low mobility factor or rate of change of zones stands a better chance, should have stable one hop links with neighbor nodes and finally should have high residual energy. Mobility is a factor that impacts the network severely. A node with high mobility will bring variations in the system criteria like distance, node density and node’s residual energy as the node will have to regulate its transmission power according to its distance with the CH or the BS. It is therefore necessary to study node mobility models that can give the relationship between the speed of nodes and the respective speed correlation

Page 22 of 36

Hence, the final expression of the fitness function selected for the proposed algorithm is fitscore = (ME + CHFrac + CHSpeed)

(19)

cr

ip t

The solution set is created with an aim of minimizing the fitscore value. Each mentioned objective has to be minimized to an acceptable level without dominance of any of the objectives over the other. The aim of using GA is to further select the CHs in such a way that the network consumes a minimum amount of energy as possible. Other steps, such as the routing of data packets, are performed in accordance with the routing of the ZEEP algorithm.

us

7. Simulation environment

an

MATLAB and network simulator version 2 have been used to test the proposed algorithm. We give the simulation parameter in Table 4 and Table 5. Section 7.1 gives the mobility model selected for the algorithm. 7.1. Mobility model

The mobility model selected is Random Waypoint. The nodes are randomly distributed, and each node selects a destination and moves toward it independent of other nodes. The node velocity is uniform but randomly chosen between [0,Vmax], where Vmax for the proposed model is 20 m/s. The nodes after reaching the destination can either stay there for some time depending on the pause time Tp or choose to move on to a new destination and repeat the whole process. For the proposed algorithm, we have chosen a Tp= 0, which means continuous mobility. Also a 0m/s speed defines a node which is immobile or stationary; however, it does not mean that this node cannot become mobile later in the course of experiments. With an advent of easily integrable low cost MEMS accelerometers it is now easy to make sensor nodes aware of their speed of movement. This is called mobility awareness and our proposed algorithm OZEEP performs clustering based on the mobility awareness of nodes to create stable clusters leading to an energy efficient data routing. The selected CH will have a mobility of less than the average nodal speed given by 0.5Vmax.

Ac ce

pt

ed

M

between them and ideally try to minimize the distance between these mobile nodes. Mobility in nodes is designed for a three dimensional topology (x, y, z). However, during design and simulation, node movement is allowed in a flat territory with coordinate value of z=0. Thus only the coordinates x and y of the node are continuously tuned with z=0 kept constant. Movement in nodes can be induced either randomly where a node starts from a random position and can move to a destination with a speed randomly generated or start or end point of the node can be fixed. In both cases whenever the position of the node is to be determined, then a trigger in form of a query can be generated by a neighbor node for distance or speed updates. In our proposal we use a very realistic mobility model called Random Waypoint Mobility model for nodes. Random Waypoint Mobility model is a spatial dependent mobility model, that is, it shows how two nodes are dependent in their motion. The model allows a node (a CH or a CM) to start from a random position and move in a random direction for a certain period of time called as pause time Tp. The node can again pick a random target (x,y) with parameters like speed between [0, Vmax] and a pause time Tp(More details on the technique are presented in section 7). The nodes which have speed less than the average speed given by 0.5 Vmax , are candidates for cluster heads. Slow moving cluster heads permit stable cluster configurations. Network simulator allows different scenario files with defined APIs to support node mobility designs. Also sensor nodes can be made mobility aware, that is, they can easily be made aware of their speeds by integrating in them, the low cost MEMS (Micro Electro Mechanical Systems) accelerometers [24]. We now define CHSpeed as the summation of the speed of all of the cluster heads in the network. Mathematically, it is represented as CHSpeed=

(18)

The variable CHSpeed helps genetic algorithm to decide between multiple CH candidates to select the ones which give the lowest speeds. In other words, for an optimal score it is important to keep the value of CHSpeed variable low. This means that the cluster heads selected are slower in movement in comparison to other candidates and therefore will give a stable cluster configuration. A large value of CHSpeed means the nodes which are the elements of CHSpeed have high mobility and therefore, if they are cluster heads then clusters would be prone to disruptions and might be possible that clusters are left without any cluster heads in course of simulation. Hence, in our proposed algorithm even if fast moving nodes participate for CH selection, the genetic algorithm in OZEEP favors slow moving nodes as cluster heads. Nodes that are stationary will have a CHSpeed component of zero.

Table 4.Simulation parameters for MATLAB Number of nodes

30-100

Area

1200 m x 1200 m

Base station/ sink

Center of the area (node ID=0), assumed stationary

Page 23 of 36

Number of iterations

100

Mobility

Continuous mobility

sensor network. Our algorithm is implemented in three steps, so we would present and discuss the results accordingly. In the following section, we view the output of the fuzzy inference system, and then, we would proceed to the next level of results obtained through GA and ns2 where we finally compare our optimized algorithm with ZEEP.

Transmission energy

50 nJ/bit

8.1 Fuzzy Logic System

Receiving energy

50 nJ/bit/m2

Multipath consumption

0.0013 pJ/bit/m4

Maximum velocity (Vmax)

Node

Minimum velocity (Vmin)

Node

20 m/s

50 nJ

Initial node energy

1J

CMRmax

15 m

CHRmax

100√2m

Crossover

0.8

Mutation

0.3

Selection

Roulette Wheel

cr

Data Aggregation

The fuzzy logic module performs the critical task of identifying nodes and nominating them for the role of CH during clustering. This task is complex as it requires the system to reduce the fuzziness in the input variables so that they can identify the thin demarcation lines that exist between these inputs to give a clean crisp output. This extraction of a crisp output from the fuzzy input requires careful decision making in representing the linguistic variables to model the specified problem. This idea leads us to study the starting point or the basic block of the fuzzy set theory, that is, to focus on the selection of the membership function to have an optimal mapping of the input and output variables. Although different applications require different representations of mapping their linguistic variables to capture the fuzziness of the set, the choice of a correct membership function is important as the shape of the MF has an effect on the fuzzy inference system. Studies indicate that to arrive at a particular MF, it is necessary for the problem to be able to break the 0-1 modeling, and consequently, it is observed that a triangular MF is capable of doing so. Additionally, if the modeling necessitates that the modal value of the fuzzy set should be a core containing only one element of the Universe of discourse, it is again preferred that a triangular MF is used. Another advantage of using a triangular MF is that only a small amount of data is needed to define the MF. Triangular MFs support hyper surfaces containing linear segments and also allow modification of parameters or modal values depending on the measured input/output of the system. They are successful in providing a good mapping of the input and output variables. Most importantly, triangular MFs satisfy the partition of unity; that is, the sum of each element x in the membership grade amounts to 1, which is generally difficult to achieve in MFs with curves. However, triangular MFs are not continuously differentiable, and therefore, for complex applications, we need special MFs (for example, problems in Quantum Mechanics need special MFs to arrive at quality solutions). In our proposal, we use a triangular MF to describe the fuzziness in our selected variables. Based on the minimum difference between the acceptable output and the observed output, we prove that although our choice of selecting a triangular MF is intuition based, it is sufficient as it gives the desired results. We compare a triangular MF with a generalized bell-

us

energy

ip t

0 m/s

Area Base station/ sink

1200 m x 1200 m Center of the area (node ID=0), assumed stationary

Maximum Node velocity (Vmax)

20 m/s

Minimum Node velocity (Vmin)

0 m/s

Pause time (Tp)

0 m/s

Simulation Time

100 s

MAC

IEEE 802.11

M

30-100

Ac ce

pt

ed

Number of nodes

an

Table 5 Simulation parameters for ns2

Mobility

Continuous mobility

Radio Range

250 m

Packet length

40 bytes

Data interval

Routing Protocol

0.25 s Zone-based energy efficient routing protocol (ZEEP)

Transmission energy

0.9 W

Receiving energy

0.8 W

Initial node energy

100 J

Mobility model

Random Way Point

8. Results and Discussions The core intent of the simulation is to optimize the energy consumption based on the energy, distance, node density and mobility of nodes forming the

Page 24 of 36

#2

Node ID = {4}, count = 1

Node IDs = {3,5,7,9,13, 16,19,20,21,2 3,30,31}, count = 12

Node ID = {10, 2}, count = 2

Node IDs = {5,7,9,10,11, 12,16,23,27,3 0, 31}, count = 11

ip t

#3

Node IDs = {1,2,3,4,5,6,7,8,9, 10,11,12,13,14,15, 16,17,19,20,21,23, 24,25,26,27,28,29, 30,31,32,33}, count = 29 Node IDs = {2,3,4,5,6,7,8,9,10 ,11,12,13,14,15,16 ,17,19,20,21,23,24 ,25,26,27,28,29,30 ,31,32}, count = 28

an

us

cr

The graphs are obtained through experimental runs, and an in-depth study of them further justifies the selection of a triangular MF. To make the experiments factual, we proceeded with clustering using fuzzy logic alone. We performed clustering with all three mentioned MFs to see their response to energy consumption. From figure 11, we can see that the network energy for a generalized bell-shaped MF diminishes faster and reaches a 0 value well before the 10 iterations are completed. This is because all nodes are performing the task of data aggregation and data routing, and therefore, they expend all of their energy in transmitting and receiving information. We have also tried to see and understand the behavior of a single node in the network, which becomes a CH in some rounds and plays the role of CM in other rounds. We study the energy usage of this arbitrary selected node for cases when it is a CH and cases when it is a CM. In our presented results, we selected the node (id= 10). In figure 12, when node 10 is a CH, we can see the rapid decline in energy. Every node starts with an initial energy of 1Joule. The y-axis gives the residual energy against the number of iterations on the x-axis. The fall in energy is obvious as the load on the CH is high. Because CH node 10 is mostly in active mode, its energy consumption is high, leading to its premature dismissal. However, in figure 13, when the same node is selected as a cluster member in some consequent rounds, its energy dissipation is reduced, as its only task is to sense and provide sensed data to the CH. However, because the number of CMs is less and they have to cater to a large number of CH, node 10’s activity is not reduced, and hence, it is not able to survive longer than 10 iterations.

Ac ce

pt

ed

M

shaped MF and trapezoidal MF to further substantiate our study. We know that a CH expends maximum energy as it is responsible for data aggregation and data dissemination for its cluster as well as for the entire network. This means that the number of cluster heads selected should be an optimal value. If a number generated is too large, maximum nodes would be CHs, and therefore, more nodes would be participating in the data communication in the network. Similarly, if the number is too low, only a few clusters would be formed. Most of the routing would use all of the CHs in their routes dispending their energy and setting off re-clustering more often. It is also possible that the clusters may not contain all of the nodes in the network, especially the nodes that are spread far apart. Far away, nodes might not fall in the range of the CH, giving rise to network partition. Some nodes present at the cluster boundaries might have to expend extra energy to transmit their data to the CH because of the distances involved between them. All of these factors lead to a non-uniform energy expenditure and thus reduce the network lifetime. We therefore present results from generating an optimal number of CH in the network in a manner that the network energy is conserved and a good CH to CM ratio is preserved. Table 6 is an extraction of a few simulation runs from using generalized bell-shaped, trapezoidal and triangular MFs in searching for an optimal number of CHs. For simulation, we have taken a small number of nodes so that we are able to check the CH fraction required to give positive results on the energy consumption of the network. From Table 6, we can see that a generalized bell-shaped curve generates a very large number of CH candidates. However, when trapezoidal MF is used, we can see that the number of candidates generated is very small, a number that is not enough to cover the entire network. The triangular MF, however, performs better. We can see that for 33 nodes, the triangular MF generates 11, 13, 11 CHs. That is, for 22 CMs, there are 11CHs,whereas, in the case of the generalized bell-shaped MF, 29 CHs are created out of 33 nodes, and only 4 CMs are present. Another extreme behavior we can see is in the trapezoidal MF, which generates 2 CHs for 33 nodes and the remaining 31 nodes are CMs. Table 6 The number of cluster heads generated in different membership functions Iterations #1

Generalized bell shaped MF Node IDs = {1,2,3,4,5,6,7,8,9, 10,11,12,13,14,15, 16,17,19,20,21,23, 24,25,26,27,28,29, 30,31,32,33}, count = 29

Trapezoidal MF Node IDs = {10,32} count = 2

Triangular MF Node IDs = {4, 7, 9,10, 17,19,23,26, 27,30,32} count = 11

Page 25 of 36

ip t

cr

ed

M

an

us

Fig.11 Total network energy using generalized bellshaped MF

with energy, distance, node density and mobility parameters, we again see that it performs poorly. During our experimentations, we observed two extreme behaviors of nodes. We present both the outcomes to understand the inconsistencies associated with using a trapezoidal MF. The drop in energy is rapid. We describe two scenarios. In scenario 1 (figure 16), we see that when node 10 is a CM, it reduces its energy at a very fast pace. We can observe an energy drop from 0.5J to 0.1J in the second iteration itself and a value of 0 when it is close to the fourth iteration. However, from figure 15 it is clear that node 10 as a cluster head performs slightly better when compared with its role as a CM. It shows a slow deterioration in the energy from 0.38J to a value of 0 until it is sustained in the network for up to 31 iterations. In another case, that is, scenario 2 from figure 19, we can see that node 10 continues in the network for 58 iterations when it is a CM and for 10 iterations when it is a CH, as seen in figure 18. This inconsistent behavior can be explained with respect to the distance of node 10 from CH in scenario 1, which leads to an extra energy dispensation by node 10 to transmit to CH and thus a fast decrease in its energy when it is a CM. Node 10 is able to live longer as a CH because of its distance to BS and low mobility with high node density. As mentioned earlier, the selection of a CH is based on these mentioned parameters; therefore, in all favorable conditions, node 10 expends less energy as a CH.

Ac ce

pt

Fig.12 Node 10 as cluster head using generalized bellshaped MF

Fig.14 Scenario 1: Total network energy using trapezoidal MF Fig.13 Node 10 as cluster member using generalized bell-shaped MF We can see an extremity in the behavior of the trapezoidal MF. From Table 6, it is clear that the number of CHs generated using the trapezoidal MF is far too low to handle all of the nodes in the network. If we perform clustering using the trapezoidal MF

Page 26 of 36

ip t cr

Fig.17 Scenario 2: Total network energy using the trapezoidal MF

ed

M

an

us

Fig.15 Scenario 1: Node 10 as a cluster head using a trapezoidal MF

Fig.18 Scenario 2: Node 10 as a cluster head using a trapezoidal MF

pt

Fig.16 Scenario 1: Node 10 as a cluster member using a trapezoidal MF

Ac ce

As against scenario 1, scenario 2 is a reverse situation. The node lives longer as a CM but degrades faster when it is a CH. This unpredictable behavior of the node could be explained again with the distance parameter of CM with the CH. Being closer to the CH allows node 10 to dissipate less energy in data reporting. In both the scenarios, the total network energy remains low as the fall is steep. Figures 14 and 17 depict the total network energy obtained when the trapezoidal MF is used.

Fig.19 Scenario 2: Node 10 as a cluster member using a trapezoidal MF We now present the results of using a triangular MF. In the graphs of triangular MF, we can clearly see from figure 20 that the network energy or the energy of a node is higher than the other two

Page 27 of 36

ip t

an

us

cr

Fig.20 Total network energy using a triangular MF

Fig.21 Node 10 as cluster head using a triangular MF

Ac ce

pt

ed

M

categories described before. The selection of CH is optimal enough to manage the random deployment and movement of nodes. The mobility factor that dominates the network and causes erratic variations in the readings of network parameters such as distance and node density and finally impacts the network energy has less dominance in this case. The CH’s distribution and number selected is optimal enough to handle every single node in the network. This means that every node is part of a cluster, with a distance that does not allow nodes to dissipate extra energy in data communication. This explains in figure 22, why node 10 sustains up to 98 iterations in the network when it is a CM. Again, from figure 21, we can see that node 10 also performs better as a CH in terms of energy usage. Because the CH and CMs are consistent in their behavior, the network energy also shows an improved performance over the other two techniques. The achieved network lifetime using a triangular MF is longer than the generalized bellshaped curve or the trapezoidal MF. We also show in figure 23 a sample of decision rules used for the triangular MF and the corresponding surface generated to bring out the relation between the four parameters energy, distance, node density and the mobility factor of a node. The aim of presenting the surface view is to show the impact of the mobility parameter on the distance, density and energy of a node. Figures 24 – 29 present these relationships. The distance parameter varies according to the movement of a node, which also in turn changes the number of surrounding nodes. In other words, the larger the distance, the greater is the need to expend extra energy to reach out to the BS. Movement of the node as well as its distance affects the energy usage of the node. In addition, according to the mobility factor, the node density would vary, making it unfavorable to be a CH. Thus, we can say that the mobility of a node is an important criterion to choose a suitable CH candidate. The curve in the graphs for node 10 as a CH or CM is because a result of its activities during a particular session. However, activities of nodes reflect their energy usage in completing a task and always show degradation in energy.

Fig.22 Node 10 as a cluster member triangular MF

Page 28 of 36

ip t

Fig.23 Sample of decision rule using triangular MF during cluster head selection

Surface

Distance

and

Energy

Fig.27 Control Surface for Mobility and Distance parameter

Ac ce

pt

ed

Fig.24 Control parameter

M

an

us

cr

Fig.26 Control Surface for Density and Distance parameter

Fig.25 Control Surface for Density and Energy parameter

Fig.28 Control Surface for Mobility and Density parameter

Page 29 of 36

Table 7 is an optimized network generated through the genetic algorithm in MATLAB. The table is the final list of nodes, their IDs and their roles in the network after their deployment. The table below is a sample of output generated using the GA. Table 7 Optimized network generated by using the genetic algorithm

Ac ce

pt

ed

M

Through our experimentations, we could see that the performance of a fuzzy system depends completely on the production of fuzzy decision rules and the membership functions proposed for each fuzzy set. However, it is seen that with an increase in the number of variables, the number of rules also increase exponentially, making it difficult for designing complete sets of rules for a larger space and hence affecting the system performance. If we see the optimal quality and optimal number of CH selection as a search problem, using the fuzzy system alone has the following effect on the hyper surface of the space. The hyper surface is the performance of a system given some performance criteria. The hyper surface obtained through a fuzzy search will have the following properties [28]. 1. The hyper surface is infinitely large as the number of possible fuzzy sets for each variable is unbounded. 2. The hyper surface seen is multimodal because different fuzzy rule sets and membership functions will have related performance. 3. Finally, we can say that the hyper surface is deceptive as similar fuzzy rule sets and membership functions may have quite different performances. Thus, considering the characteristics of GA, it is observed that by using the genetic algorithm, it is possible to search the high-dimensional space and give the optimal location of this hyper surface easily. However, by using the fuzzy system, the GA is relieved of searching a larger search space as the hunt is now limited only to the nominated candidates. The search of optimal choice of CH becomes speedy as the space of the search is smaller compared to searching through all of the nodes. In addition, by using the GA, we are able to generate a balanced clustering in a manner that the network lifetime is increased.

Role Member Member Member Head Member Member Head Member Head Head Member Member Member Member Member Member Member Member Head Member Member Member Head Member Member Head Member Member Member Head Member Head Member

ip t

408 244 548 429 473 383 119 187 336 224 401 666 311 490 682 348 500 476 299 134 688 441 179 278 310 157 300 338 453 427 210 424 378

cr

680 487 552 823 808 158 391 183 729 752 40 320 973 55 901 89 636 778 786 635 90 326 1031 230 1001 314 443 166 125 41 852 486 93

Y coordinate

us

8.2 Genetic Algorithm

X coordinate

an

Fig.29 Control Surface for Mobility and Energy parameter

Node ID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33

This output obtained after a GA search is fed into the network simulator, which in turn deploys the network in ns2 and runs the ZEEP routing on the nodes. We finally compare the results of this optimized network OZEEP with ZEEP. Figure 30 is a visual of optimized network deployment done by ns2, with inputs taken from a GA search for CHs. The base station (ID=0) is centered on the defined area and is assumed to be stationary.

Page 30 of 36

ip t

cr

Fig.31 ZEEP versus OZEEP – Packet Drop

us

Fig.30 Network simulator animator showing node positions in the network generated by using genetic algorithm

Ac ce

pt

ed

M

The optimized network generated from GA is given as an input to the network simulator. The simulator generates the network and assigns each node its role. The GA generates a cluster set that is balanced and well distributed so that the network energy consumption is minimized. The results obtained show that OZEEP has almost zero packet drop (figure 31). This is because we feed an optimized clustered network, and therefore, no clustering happens during the simulation process; hence, no packet drops occur. Routing starts as soon as there is data to be transmitted. However, the scenario is not the same for ZEEP. When the simulations start, ZEEP first performs clustering and then initiates the routing protocol. This results in packet drops in the network. In addition, from figure 32, we can see that the results show a good improvement in the network lifetime when OZEEP is used as against ZEEP. Table 8 gives the percentage improvement of OZEEP over ZEEP in terms of the remaining energy during different activity progressions in the network. An average increase of 15.47% increments in the network lifetime is seen with OZEEP. OZEEP is able to conserve a node’s energy due to balanced cluster formation.

an

8.3 ns2

Fig.32 ZEEP versus OZEEP- Energy usage to reflect network lifetime

Table 8 Percentage improvement in network lifetime of OZEEP over ZEEP Energy remaining (Joules)

Time (seconds) for ZEEP

Time (seconds) for OZEEP

% improvement in network lifetime of OZEEP over ZEEP 7.14%

85

1.44

1.56

84

1.56

1.72

9.68%

82

1.83

2.06

10.81%

80

2.11

2.42

12.64%

78

2.44

2.83

13.73%

76

2.72

3.11

12.50%

74

3.00

3.50

14.29%

72

3.28

3.83

14.49%

70

3.56

4.14

14.09%

68

3.83

4.56

15.85%

66

4.06

4.94

17.98%

64

4.28

5.28

18.95%

62

4.56

5.67

19.61%

60

4.83

6.00

19.44%

58

5.08

6.36

20.09%

56

5.33

6.72

20.66%

Page 31 of 36

8.4 Energy Overhead analysis

Ac ce

pt

ed

M

In sensor networks energy is the most crucial parameter and thus it is important to devise optimal methods to use this resource efficiently so that the network lifetime is increased. Authors have proposed techniques in literature to efficiently utilize node’s energy so that network lifetime could be incremented. As part of our work we also propose an energy efficient cluster based routing protocol called OZEEP. We have tried to optimize the cluster head selection technique of existing Zone based energy efficient routing protocol for mobile nodes (ZEEP). To achieve this we use a genetic fuzzy system with defined system parameters to select the best node among the N randomly deployed nodes to be cluster heads. Once clustering is achieved the underlying routing protocol used is ZEEP. In ZEEP each node is aware of its location information and the location of the sink. Whenever a data request arrives, ZEEP performs greedy forwarding of the data based on the next hop neighbor’s location information and its distance from the base station. However, ZEEP sends a control packet towards the next hop node to check whether route exists to the sink or not. This is done prior to transmitting the data packet. Considering best case scenario where all links are good and route exists to the base station: if the sink is 3 hops away then the number of control packets generated for route existence is 3 and one data packet. Size of a control packet is 32 bits and that of the data packet is 40 bytes. Therefore, the best case is the shortest path selected towards the sink (assuming all nodes in the path lead to the sink). The worst case scenario would be the time spent in identifying the node that can be on the intermediate path for reaching the destination and the path could be the longest route or the diameter of the network to be traversed. Since no route set up is required to be maintained for any data transmission in ZEEP and because of its dynamic forwarding methodology, ZEEP can be seen as similar to greedy forwarding where link or path breakage due to mobility of nodes do not have much impact on the data routing. Bandwidth utilization of ZEEP is efficient as it does not generate route establishment messages for the network. It does not overload the network with control packets. Hence, approving ZEEP as an underlying protocol, results in a flawless data routing. However, ZEEP’s limitation of selecting cluster or zone heads based on mobility factor does not hold good for some scenarios where two nodes have similar values for the mobility factor. Therefore, we have tried to optimize the cluster head selection technique of ZEEP by using a GFS based system. In the cluster set up phase the fuzzy module in the nodes generate node’s status to be a cluster head for that round. Energy and message overhead would be extremely high when all N nodes in the

network declare themselves as CH candidates. The maximum (dominating) independent set for CH candidates would have a cardinality N and therefore N. ETx transmissions will happen in the network. The network energy consumed during these transmissions will depend on the distance of nodes with the base station and would be high for nodes farther away. Also the number of messages generated would be N. However, genetic algorithm will generate only the optimal number of candidates to be CH, hence a single message transmission alone from the base station would be enough to inform the network about the selected nodes as CH. The advertisement phase of informing the nodes to join a cluster will have multiple message exchanges between the nodes and the cluster heads. The number of messages generated would be bounded by the number of one hop neighbors of the CH. For example, a single message from CH could be responded by 10 one hop away nodes in its vicinity. That means 10 messages from the non CH nodes to CH would be generated. However, since the distance for intra cluster communication is less, therefore, the amount energy required for transmission between CH and non CHs would be lower when compared to the inter cluster transmissions that will happen between cluster heads. Thus, the total number of message exchanges during set up phase will depend on the number of CHs selected for that round and will vary in subsequent rounds. Once clustering is completed and an optimal clustered network configuration is obtained, then the data routing phase starts. For a network of (area = A, number of nodes = N, communication radius = r and sink location) with clustering scenario (energy, mobility, distance, density, CH speed, CH fraction required for CH to CM ratio, and mean communication energy), the communication overhead is the sum of sent and received packets from all the nodes in one round. If we consider one round then each CM sends one message to the CH and the CH sends one packet to the base station. The number of messages in the network will depend on the ratio of CHs to CMs, that is, total number of clusters, CHs and CMs. The energy consumed is mainly because of the amount of intra and inter cluster communications. Further, sensor nodes adopt different operational modes for energy conservation. The sensor nodes are typically in an idle mode until some event occurs. This period could be longer than expected and keeping nodes in a switched off mode might lead to a miss in the event detection, therefore, it is sensible to put the nodes in the sleep mode. The nodes can wake up and made active when an event is triggered or the time expires. The active node can now take necessary actions (sense, store, report and aggregate) on the event that triggered its change of state. Sometimes nodes can be allowed to take a deeper sleep mode by switching off the radio front end sensor and the memory component to save more energy. Though

ip t

20.97%

cr

6.89

us

5.44

an

55

Page 32 of 36

8.5 Complexity analysis

ip t

Two measures of complexity are usually considered for synchronous algorithms: time complexity and communication complexity. We define time complexity as the number of rounds or iterations an algorithm takes to output a desired result or until the algorithm halts leading to no more change in the outputs from further iterations. For our proposed algorithm, with each n-bit binary chromosome representing a cluster arrangement, there can be 2n possible configurations. The running time of OZEEP’s fitness function with a cluster configuration as an input is polynomial in n. An extensive search for an optimal solution would require 2n summons of the fitness function. With OZEEP, the maximum number of invocations is reduced to P x G (where constants P and G represent the population size and number of generations respectively), to achieve near optimal solution. The communication complexity is typically measured in terms of the total number of non-null messages that are sent or successfully delivered. Occasionally, we also take into account the number of bits in the message. The communication complexity is mainly significant if it causes enough congestion to slow down processing. For our proposed algorithm OZEEP, benefit of using GFS based system for CH selection is that, fuzzy module allows only CH contenders to communicate with the base station thus reducing the number of nodes reporting to the base station and therefore avoids collision among nodes. Further, the clustering operation is done by the base station which outputs the ids of nodes selected as cluster heads to the entire network. This avoids any kind of unnecessary messages in the network leading to successful delivery of messages generated in the network. The packet drop ratio of OZEEP is zero which proves that the implementation phase of OZEEP allows all generated messages to be delivered successfully. This is achieved because of optimal clustering done using GFS based system.

M

Fig.33 Energy savings and overheads for sleep modes

an

us

cr

these multiple states of operations in sensor nodes avoid unnecessary energy usage however, they still do use some energy during mode transitions giving rise to energy operational overheads. The deeper the sleep, the more is the energy spent during waking up the components. Figure 33 below shows this idea of changing modes of operations from active to sleep to active mode again. The diagram shows the energy overhead generated for mode transitions. This value might be low but it is an important factor to be considered to perform energy analysis as the network consists of sensor nodes explicitly known for being energy constraint.

Ac ce

pt

ed

It is important to decide when a node should go in to sleep and when it should be activated. From the figure we can see that at time t1 the decision to change the operation mode of node from Pactive to Psleep is taken. It takes τdown time to reach the sleep state. If the node remains in the active state and an event occurs at tevent then the amount of energy spent during idle period is Eactive = Pactive (tevent – t1). However, if the node is in sleep mode then the power consumption is τdown (Pactive + Psleep)/2 + (tevent - t1 τdown) Psleep, assuming (Pactive + Psleep)/2 is average power consumption in this phase and Psleep is the power consumed until tevent. The energy saving is given by Esaved = Pactive (tevent – t1) - τdown (Pactive + Psleep)/2 + (tevent - t1 - τdown) Psleep (20) The energy overhead is given by– Eoverhead = τup (Pactive + Psleep)/2

(21) This Energy overhead is certainly an overhead as no other activity can be executed during this period. A sleep mode should be permitted only if Eoverhead < Esaved. The different scenarios of functioning of OZEEP along with the energy constraints imposed by sensor nodes themselves make the energy overhead analysis of such a complex environment very extensive.

9. Conclusions In this work, we have used a GFS system to perform clustering over our existing routing algorithm, ZEEP. The parameters selected as criterion for CH selection are energy, distance, node density and mobility of a node. We can say that mobility is the main criterion that dominates the behavior of all of the other parameters. Mobility of a node defines the variation in the constraints such as distance and energy, and consequently, these correlated issues affect the node density. However, by using a twophase CH selection with a fuzzy system identifying the optimal candidates and GA further selecting the

Page 33 of 36

ip t

cr

M

The future endeavor is to extend OZEEP by introducing caching in the network. We plan to study the data access delay of OZEEP and design caching in a way to reduce the latency. The aim will be to further reduce network-wide transmissions leading to an increased network lifetime.

us

10. Future Works

Logic, TENCON 2007-2007 IEEE Region 10 Conference, pages 14. DOI: 10.109/TENCON.2007.4428982 [9] J. R. Srivastava, TSB. Sudarshan, ZEEP: Zone based Energy Efficient Routing Protocol for Mobile Sensor Networks, IEEE International Conference on Advances in Computing, Communications and Informatics (ICACCI), August 2013 [10] J. Jhang, Y. Lin, C. Zhou, J. Ouyang, Optimal Model for Energy-Efficient Clustering in Wireless Sensor Networks Using Global Simulated Annealing Genetic Algorithm, International Symposium on Intelligent Information Technology Application Workshops, IEEE 2008 [11] L.B. Leal, R. A. L. Rabelo, R. H. Filho, F.A.S. Borges, An Application of on Genetic Fuzzy Systems for Wireless Sensor Networks, IEEE International Conference on Fuzzy Systems, Taiwan, June 27-31, 2011 [12] M. Karimi, H. R. Naji, S. Golestani, Optimizing Cluster-Head selection in Wireless Sensor Networks using Genetic Algorithm and Harmony Search Algorithm, 20th Iranian Conference on Electrical Engineering, (ICEE2012), May 15-17,2012, Tehran, Iran, IEEE 2012 [13] M. M. Afsar, Mohammad-H, Tayarani-N, Clustering in Sensor Networks: A literature survey, Journal of Network and Computer Applications 46(2014) 198-226, Elsevier [14] M. Omari, H. Abdelkarim, B. Salem, Global and Local sensor clustering using Genetic Algorithms and Fuzzy Logic, CTIC 2012, University d Adrar, Algerie [15] M. Song, Z. Cheng-lin, Z. Zhou, Y. Ye, An Improved Fuzzy Unequal Clustering Algorithm for Wireless Sensor Network, Mobile NetwAppl (2013) 18:206-214 DOI 10.1007/s11036-0120356-4 [16] O. Cordon, F. Gomide, F. Herrera, F. Hoffmann, L. Magdalena, Ten years of genetic fuzzy system: current framework and new trends, Journal of Fussy sets and System, 141(2004)5-31, Elsevier [17] P. V. Vinayak, S. Sawarkar, A. Gawande, Efficient Energy Consumption In Two Tiered Sensor Networks Using Genetic Algorithm, ARTCOM’10 Proceedings of the 2010 International Conference on Advances in Recent Technologies in Communication and Computing, pages 295-300, 2010. DOI: 10.1109/ARTCOM.2010.89 [18] Q. Y. Zhang, Z-M Sun, F. Zhang, A Clustering Routing Protocol for Wireless Sensor Networks Based on Type-2 Fuzzy Logic and ACO, IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), July 6-11, China, 2014 [19] R. Arya, S.C. Sharma, Proceedings of the Third International Conference on Soft Computing for Problem Solving, Advances in Intelligent Systems and Computing 259, DOI: 10.1007/978-81322-1768-8_19, Springer India 2014 [20] R. Khoshkangini, S. Zaboli, S. Sampalli, Energy Efficient Clustering using Fuzzy Logic, International Journal of Computer Science and Mobile Computing, ICMIC13, December 2013, pg 814. [21] S. Babaie , S. Shokraneh , A. Ghaffari, A. Jahangiry, CCGA: Clustering based on Cluster head with Genetic Algorithm in Wireless Sensor Network, IEEE 2010 International Conference on Computational Intelligence and Communication Systems, pages 367-371, 2010. DOI: 10.1109/CICN.2010.150 [22] S. E. Hashemi, H. Motameni, M.R. Ghaleh, S. Esmaeili, Clustering and routing wireless sensor network based on the parameters of distance, density, energy and traffic With the help of fuzzy logic, IJCSI International Journal of Computer Science Issues, Vol. 10, Issue 3, No 2, May 2013, ISSN (Print): 1694-0814| ISSN (Online): 1694-0784 [23] S. M. Abedini, A. Karimi, Increasing the lifetime of heterogenous sensor network by using Genetic-fuzzy Clustering, International Journal of Smart Sensors and Ad-Hoc Networks (IJSSAN), ISSN No: 2248-9788(Print), Vol-2, ISSN 1,2, 2012 [24] S. Sarangi, S. Kar, Genetic Algorithm based Mobility Aware Clustering for Energy Efficient Routing in Wireless Sensor Networks, Proceedings of the 17th IEEE International conference on Networks, ICON 2011, Singapore, December 14-16, 011. IEEE 2011 ISBN 978-1-4577-1824-3 [25] S. Wazed, A. Bari, A. Jaekel, S. Bandyopadhyay, Genetic Algorithm Based Approach for Extending the Lifetime of Two-

an

best amongst them, we are able to achieve a balanced clustering such that the network lifetime is increased. Since, OZEEP is an optimized network which is used for data routing, therefore, the need for retransmissions of data and route setups in case of path breaks are completely missing. Thus, the total energy consumption of the network while routing of packets is also greatly reduced as can be seen from the experimental results. Our proposed optimized algorithm OZEEP is observed to perform better over ZEEP. Also, the proposed algorithm is implementable on real sensor nodes as the main computing, that is, running of genetic algorithm, which requires multiple iterations to converge to a desirable result is done by the base station or SINK node, assumed to be a processor with unlimited power and energy.

References

ed

Conflict of Interest The authors declare that they have no conflict of interest.

Ac ce

pt

[1] A.A.Abbasi, M. Younis, A survey on clustering algorithms for wireless sensor networks, Network Coverage and Routing Schemes for Wireless Sensor Networks, Volume 30, Issues 14–15, Pages 2697-2994, Elsevier 2007 [2] A. Rahmanian, H. Omranpour, M. Akbari, K. Raahemifar, A Novel Genetic Algorithm in LEACH-C Routing Protocol for Sensor Networks, 24th Canadian Conference on Electrical and Computer Engineering (CCECE), pages 001096-001100, C, IEEE 2011. DOI: 10.1109/CCECE.2011.6030631 [3] E. Saeedian, M. Jalali, M. M. Tajari, M. Torshiz, G. Taday on CFGA : Clustering wireless sensor network using fuzzy logic and genetic algorithm”, WiCom - International Conference on Wireless Communications, Networking and Mobile Computing, 1-4, 2011 [4] F. M. Filho, F. Gomide, Fuzzy Clustering in Fitness Estimation Models for Genetic Algorithms and Applications, IEEE International Conference on Fuzzy Systems, Vancover, Canada, July 16-21, 2006 [5] H. Bagci, A. Yazici, An energy aware fuzzy approach to unequal clustering in wireless sensor networks, Journal of Soft Computing, 13(2013), 1741-1749 [6] H-S .Seo, S-J. Oh, C-W. Lee, Evolutionary Genetic Algorithm for Efficient Clustering of Wireless Sensor Networks, Proceedings of the 6th IEEE Conference on Consumer Communications and Networking Conference, Pages 258-262, 2009 [7] H. Taheri, P. Neamatollahi, O. M. Younis, S. Naghibzadeh, M. H. Yaghmaee, An energy-aware distributed clustering protocol in wireless sensor networks using fuzzy logic, Journal of Ad-Hoc Networks, 2012 [8] J. Anno, L. Barolli, F. Xhafa, A. Durresi, A Cluster Head Selection Method for Wireless Sensor Networks based on Fuzzy

Page 34 of 36

Computer Science and Information Science at BITS, Pilani, India. He is an active member of IEEE USA, ACM USA, International Association of Engineers (IAENG), USA, Technical Advisory Board, Cradle Technologies, U SA. He is Life member of ISTE, India and fellow member at Association of Computer Electronics and Electrical Engineers (ACEEE). He is reviewer of several international conferences and journals and has many research publications to his credit. His research interests include computer architecture, memory design for embedded systems, reconfigurable system design, multicore architectures, routing & caching in sensor networks, and robotics. He is currently serving as a Professor & Chairman of the Department of Computer Science and Engineering at Amrita VishwaVidyapeetham University, School of Engineering, Bangalore Campus, India.

Ac ce

pt

ed

M

Ms. Juhi R Srivastava is a full time Ph.D. scholar with Amrita VishwaVidyapeetham University. She has been associated as a faculty member with Department of Computer Science and Engineering, Amrita School of Engineering, Bangalore campus for more than 8 years (2005-2013). She was serving as Assistant Professor, Sr. Grade before taking full time research in 2013. She holds an M.Tech degree in Scientific Computing (2004) and Bachelor of Engineering in Electronics and Communications (2002), both from Birla Institute of Technology, Deemed University, Mesra, Ranchi, India. Her research interest includes soft computing techniques, data communication and networking, wireless networks, ad-hoc sensor networks and distributed algorithm. She is a student member IEEE. Her published research details available at http://orcid.org/0000-0001-5226-5383

an

us

cr

ip t

Tiered Sensor Network, ,IEEE International Symposium on Wireless Pervasive Computing, 2007. [26] V. Godbole, Performance Analysis of Clustering Protocol Using Fuzzy Logic for Wireless Sensor Network, IAES International Journal of Artificial Intelligence (IJ-AI), Vol. 1, No. 3, September 2012, pp. 103-11, ISSN 2252-8938 [27] X. Miao, H. Ting-Lei, Z. Xiao-Shu, A Novel Routing Algorithm for Energy-Efficient in Wireless Sensor Networks, IEEE Third International Conference on Genetic and Evolutionary Computing, 2009 [28] Y. Shi, R. Eberhart, Y. Chen, Implementation of Evolutionary Fuzzy Systems, IEEE Transactions on Fuzzy System, Vol. 7, No. 2, April 1999

Dr. T. S. B. Sudarshan completed his Ph. D. in Techniques to Enhance Web Performance in Fixed Networks and Mobile Networks from BITS, Pilani, India in 2007. He has also served as Head of the Department of

Page 35 of 36

Ac

ce

pt

ed

M

an

us

cr

i

Graphical abstract (for review)

Page 36 of 36