A VIRTUAL INFRASTRUCTURE FOR MITIGATING TYPICAL CHALLENGES IN SENSOR NETWORKS

Transcription

1 A VIRTUAL INFRASTRUCTURE FOR MITIGATING TYPICAL CHALLENGES IN SENSOR NETWORKS by Hdy S. Abdel Slm B.S. June 1999, Alexndri University, Egypt M.S. June 2004, Alexndri University, Egypt A Disserttion Submitted to the Fculty of Old Dominion University in Prtil Fulfillment of the Requirement for the Degree of DOCTOR OF PHILOSOPHY COMPUTER SCIENCE OLD DOMINION UNIVERSITY November 2010 Approved by: Stephn Olriu (Director) Kurt Mly Hussein Abdel-Whb Lrry Wilson Ivn Stojmenović

2 ABSTRACT A VIRTUAL INFRASTRUCTURE FOR MITIGATING TYPICAL CHALLENGES IN SENSOR NETWORKS Hdy S. Abdel Slm Old Dominion University, 2011 Director: Dr. Stephn Olriu Sensor networks hve their own distinguishing chrcteristics tht set them prt from other types of networks. Typiclly, the sensors re deployed in lrge numbers nd in rndom fshion nd the resulting sensor network is expected to self-orgnize in support of the mission for which it ws deployed. Becuse of the rndom deployment of sensors tht re often scttered from n overflying ircrft, the resulting network is not esy to mnge since the sensors do not know their loction, do not know how to ggregte their sensory dt nd where nd how to route the ggregted dt. The limited energy budget vilble to sensors mkes things much worse. To sve their energy, sensors hve to sleep nd wke up synchronously. However, while promoting energy wreness, these ctions continully chnge the underlying network topology nd mke the bsic network protocols more complex. Severl techniques hve been proposed in different res of sensor networks. Most of these techniques ttempt to solve one problem in isoltion from the others, hence protocol designers hve to fce the sme common chllenges gin nd gin. This, in turn, hs direct impct on the complexity of the proposed protocols nd on energy consumption. Insted of using this pproch we propose to construct lightweight bckbone tht cn help mitigte mny of the typicl chllenges in sensor networks nd llow the development of simpler network protocols. Our bckbone construction protocol strts by tiling the re round ech sink using identicl regulr hexgons. After tht, the closest sensor to the center of ech of these hexgons is determined we refer to these sensors s bckbone sensors. We define ternry coordinte system to refer to hexgons. The resulting system provides complete set of communiction pths tht cn be used by ny geogrphic routing technique to simplify dt communiction cross the network. We show how the constructed bckbone cn help mitigte mny of the typicl chllenges inherent to sensor networks. In ddition to sensor locliztion, the network

3 bckbone provides n implicit clustering mechnism in which ech hexgon represents cluster nd the bckbone sensor round its center represents the cluster hed. As cluster heds, bckbone sensors cn be used to coordinte tsk ssignment, workforce selection, nd dt ggregtion for different sensing tsks. They lso cn be used to loclly synchronize nd djust the duty cycle of non-bckbone sensors in their neighborhood. Finlly, we propose Bckbone Switching, technique tht cretes lterntive bckbones nd periodiclly switches between them in order to blnce energy consumption mong sensors by distributing the dditionl lod of being prt of the bckbone over lrger number of sensors.

4 c Copyright, 2011, by Hdy S. Abdel Slm, All Rights Reserved iv

5 I dedicte this thesis to the soul of my mother, My ALLAH shower ll his mercy nd blessings upon her nd grnt her plce in Jnnt Alfirdus... Amen v

6 ACKNOWLEDGMENTS This work could not be completed without the help of mny individuls to whom I would like to express my pprecition. First nd foremost, I would like to thnk my dvisor, Prof. Stephn Olriu, who hs put gret del of his time nd effort into the guidnce of this work. Prof. Olriu used to tret me s his son nd not s his student. He used to keep his office opened in front of me in order to help me solve ny technicl or personl problem I hd during the progrm. He funded me through his reserch grnts nd never stopped encourging me till I could complete my PhD. Next, I would like to convey my sincere thnks to other members of my PhD committee, Prof. Kurt Mly, Prof. Hussein Abdel-Whb, Prof. Lrry Wilson, nd Prof. Ivn Stojmenović. Their expertise, thorough reviewing, continuous support, nd vluble suggestions hve led to gretly improved disserttion. I m lso grteful to my fmily for their encourgement nd support. Finlly, specil thnks to my fther nd mother for their understnding nd ptience during the time I spend completing this work. vi

7 vii TABLE OF CONTENTS Pge LIST OF TABLES ix LIST OF FIGURES xi CHAPTERS I Introduction I.1 Design Chllenges I.2 Motivtion I.3 Sensor Network Model I.4 Rodmp II Stte Of The Art II.1 Sensor Bckbones nd Infrstructures II.2 Sensor Locliztion II.3 Sleeping Schedule III Bckbone Construction Protocol III.1 Communiction bckbones III.2 The Bckbone Construction Protocol III.2.1 Computing nd Mesuring Angles III.2.2 Order of Selection of Bckbone Sensors III.2.3 The Detils of Bckbone Selection III.3 Bckbone Switching III.4 Recovering from Sensor Voids III.4.1 Even-Neighbor Replcement the Detils III.4.2 Bckwrd Selection the Detils III.5 Mitigting Network Chllenges III.5.1 Sensor Locliztion III.5.2 Clustering nd Leder Election III.5.3 Geogrphic Routing III.5.4 Dt Aggregtion: III.6 Simultion Results IV Bckbone-Bsed Tsk Mngement IV.1 Centrlized Tsk Mngement IV.1.1 Centrlized Tsking Model IV.1.2 CTW Messges IV.1.3 Workforce Selection IV.2 Distributed Tsk Assignment IV.2.1 Phse 1: Estimting the Mximum Energy IV.2.2 Phse 2: To Prticipte or Not to Prticipte IV.2.3 Averge Over-Recruited Workforce IV.3 Dt Aggregtion IV.4 Simultion Results

8 viii V Scheduling in Sensor Networks V.1 Applicbility of PASTA V.2 Resoning bout the time-independent wke probbility V.2.1 Empiricl vlidtion V.3 Scheduling Schemes V.3.1 Sttic Scheduling V.3.2 Dynmic Scheduling V.3.3 Energy-Awre Scheduling V.4 Energy Estimtion V.5 Performnce Evlution VI Conclusions VI.1 Future Reserch Directions BIBLIOGRAPHY APPENDICES X A Distribution nd Expecttion of X+Y A.0.1 Cse 1 (d bc) A.0.2 Cse 2 (d > bc) A.1 Evluting E[Z] A.1.1 Cse 1 (d bc): A.1.2 Cse 2 (d > bc): VITA

9 ix LIST OF TABLES Pge 1 Single-One Representtions of 0 m < 2 3 = Percentge of network lifetime lived before density degrdes Mic2 Power Requirements Percentge of network lifetime lived before density degrdes

10 x LIST OF FIGURES Pge 1 Clusters constructed from rndomly selected bckbone sensors Hexgonl clusters mximize the number of bckbone neighbors Prtitioning the re round the sink into sectors Estimtion of the position of hexgon centers Rdition lobes nd bemwidths of directionl ntenn pttern Cse 1: estimtion of e s Cse 2: estimtion of e b Blncing energy consumption using bckbone switching Unloclized sensors in the shdow of void regions Illustrting even neighbor replcement Recovering from voids using even neighbor replcement Even-neighbor replcement with bckwrd selection Voids recovery using even neighbor replcement/bckwrd selection Dt ggregtion nd routing through our bckbone Averge locliztion error of bckbone sensors vs. row number for different network densities using single sink in the bsence of mesurements errors Actul hexgons produced by simultion Actul vs estimted positions of bckbone sensors Averge locliztion error in the presence of errors in distnce nd/or ngle mesurements for different network densities A comprison between verge locliztion error for different locliztion protocols using different network densities An exmple of improper tsk ssignment Minimum distnce between two concurrent tsks Tsking model for the centrlized pproch Tsking model for the distributed protocol Estimtion of the mximum energy mong cndidte sensors Dividing the sensing rnge into k disjoint subregions E[Y ] vs. k/n Averge number of CTW messges Averge number of bidding rounds Averge number of bidding slots in bidding rounds Averge width of sensor energy spectrum Avg. width of energy spectrum for different initil energy levels Averge degrde in network density throughout its 0.2-relible-lifetime using different protocols. () ρ = 0.3, (b) ρ = 0.7, (c) ρ = The growth rte of energy holes in the network 0.2-relible-lifetime when using different workforce selection protocols nd different network densities

11 34 () Totl number of executed tsks (b) Totl number of tsks executed using the exct workforce size Illustrting the renewl process of wke-up times Empiricl vlidtion of Theorem V Sleep nd wke cycles in sttic scheduling () Averge number of tsks, (b) Percentge of under-recruited tsks, (c) Averge number of relible tsks, executed in the network throughout its 0.2-relible-lifetime using different scheduling schemes Sensor energy distribution in subsequent stges of the network lifetime for different scheduling protocols The verge width of sensor energy spectrum Averge degrde in network density throughout its 0.2-relible-lifetime using different scheduling protocols. () ρ = 0.3, (b) ρ = 0.7, (c) ρ = A comprison between the growth rte of energy holes throughout 0.2- relible-lifetime of the network using different scheduling protocols for different network densities. () ρ = 0.3, (b) ρ = 0.7, (c) ρ = An illustrtion of different possible cses or different rnges of the rndom vribles X nd Y An illustrtion of the different subcses of Cse An illustrtion of the different subcses of Cse xi

12 1 CHAPTER I INTRODUCTION Electro-mechnicl sensors hve been used for reltively long time in different control systems. In typicl such system, smll number of sensors re deployed in predetermined positions in order to provide redings bout importnt system prmeters. Through wired-bsed networking of sensors, sensory dt re trnsmitted to centrl processing unit which nlyzes received dt nd tke decisions tht control the functionlity of the system. Advnces in nno-technology nd wireless communictions hve enbled the development of new genertion of sensor-bsed networks. In prticulr, technology llowed the mssive production of low-cost low-power multi-functionl sensors. Although, these sensors usully hve limited sensing, computtionl nd communiction cpbilities, they cn be networked to provide services for vst spectrum of pplictions. The pst decde hs witnessed phenomenl prolifertion of sensor network pplictions rnging from bttlefield surveillnce [1], to border monitoring [2], to fire detection nd hbitt monitoring [3], to home utomtion [4], to trffic control [5], to helth-cre [6], nd to body sensor networks [7], mong mny others. I.1 DESIGN CHALLENGES Designing relible nd energy wre sensor network hs been lwys chllenging due to mny fctors inherent to the modest sensor resources nd to the nture of the sensor network itself. A considerble mount of reserch hs been conducted, nd is still ongoing, on the topic of developing protocols nd solutions to overcome these chllenges. We now highlight some of the min chllenges specific to the design of efficient protocols for sensor networks: Ad hoc nture: sensor networks re d hoc in nture with no underlying infrstructure. It is the responsibility of individul sensors to identify their connectivity to other sensors nd to decide wht routing mechnism should be used to forwrd informtion to their intended destintion. Moreover, trditionl routing schemes my not be useful here becuse of the dynmic topology nd energy considertions; This disserttion follows the style of The Physicl Review

13 2 Limited energy budget: the sensors re powered by modest, non-renewble on-bord energy source. Once its energy is depleted, sensor becomes totlly non-functionl. Hence, sensor energy must be treted s precious resource tht hs to be used wisely, if network longevity is to be promoted; Energy hole problem: sensory dt collected from different prts of the network need to be routed towrds one of the network sinks for processing/ggregtion. The dditionl routing lod imposed on network sensors round sinks, would result in the depletion of their energy much fster thn other sensors. Once the energy of these sensors is totlly depleted, the sinks re disconnected from the rest of the network by holes tht contin only non-functionl sensors. In time, sensory dt cn not be routed to the sinks nd the network fils; Loction unwreness: the sensors re usully deployed in regions tht hve no infrstructure t ll. A typicl wy of deployment is to sctter the sensors from irplnes. This kind of deployment does not llow sensors to be wre of their positions. Moreover, ssuming tht the sensors cn be equipped with reltively expensive nd energy-hungry GPS chips does not seem to be fesible or, indeed, n cceptble ssumption for these low-cost low-power devices; The sensors must work unttended: due to the mssive deployment of sensors, it is entirely imprcticl to devote ttention to individul sensors. Once deployed, the sensors must work unttended with no externl intervention; Limited computing power: the sensors re designed to be low-cost, low-power devices. Thus, the computing cpbilities of sensors re very limited in terms of the processing speed nd vilble memory. This imposes dditionl restrictions on the type of protocols tht cn be run by sensors; Smll trnsmission rnge: wireless communiction re known to consume lrge portion of sensor energy; indeed, running the rdio interfce cuses the lrgest energy expenditure incurred by individul sensors. Supporting sensors with long-rnge trnsmission cpbility cn excessively consume their energy which it turn reduces their lifetime drmticlly. To promote the functionl longevity of the network, the sensors should perform their tsks with the minimum possible sensor-to-sensor or sensor-to-sink communiction;

14 3 Dynmic topology: to sve their energy, the sensors lternte between sleep nd wke periods. Due to clock drift nd lck of communiction, the sleep nd wke cycles (duty cycles) of different sensors re ssumed to occur synchronously. When sensor wkes up, it might find mny of its neighboring sensors still sleeping. This behvior continuously chnges the connectivity of the network creting dynmic topology tht complictes mny of the network tsks, including routing nd coverge; Unlike clssicl networks, where the min trget is to mximize chnnel throughput or link utiliztion, the min trget in sensor networks is to extend the network lifetime without scrificing coverge, connectivity, nd relibility of the network. I.2 MOTIVATION Severl techniques hve been proposed to ddress ech of the chllenges mentioned erlier (i.e. locliztion, clustering, routing, dt ggregtion, etc). The min gol of these techniques ws to mke the network more trctble by solving one of the inherent network chllenges. A mjor problem when using this pproch, ech of these techniques tries to solve one problem seprtely from other problems. Protocol designers hve to fce the sme common chllenges every time they solve ny of these problems. This in turn hs its direct impct on the complexity of the proposed protocols nd energy consumption. Insted of solving ech of these problems individully fcing the sme common chllenges with ech problem, we propose to construct wht we cll network skeleton tht is constructed immeditely fter network deployment nd provides some kind of n infrstructure tht mkes the network more trctble. The skeleton provides sensors with corse locliztion informtion tht enbles them to ssocite their sensory dt with the geogrphic loction in which the dt ws mesured. Moreover, it promotes geogrphic routing scheme tht simplifies dt communiction cross the network through skeleton sensors. By hypotheticlly tiling the deployment re using identicl hexgons, the construction lgorithm clusters sensors bsed on their loctions into hexgons(clusters). Skeleton sensors which re chosen to be the closest sensors to the centers of these hexgons represent cluster heds nd cn ply crucil rule in coordinting tsk ssignment, workforce selection, nd dt ggregtion.

15 4 I.3 SENSOR NETWORK MODEL We ssume typicl sensor network which includes mssive deployment of tiny sensors ech perhps no lrger thn dime. These sensors re deployed uniformly t rndom cross the deployment re. In ddition to the tiny sensors, we ssume the existence of more powerful devices, referred to s sinks. While the tiny sensors re sttic, the sinks my be be ble to move in support of the mission t hnd. Further, while the sensors hve modest non-renewble energy budget, the sinks re ssumed to be energeticlly self-sufficient nd/or to hve rechrgeble btteries. Finlly, while the number of sensors is lrge, perhps in the tens of thousnds, the number of sinks is mny orders of mgnitude smller. For precise reference, we now list our ssumptions bout the cpbilities of sensors nd sinks. I- Sensors: The sensors re tiny, inexpensive devices with very limited sensing, computtionl nd communiction cpbilities; The sensors re powered by non-renewble on-bord energy source; when the energy budget is exhusted the sensor becomes in-opertionl. Hence, the sensors sleep nd wke up lterntively to sve their energy. Sleep nd wke-up cycles for different sensors re ssumed to occur synchronously; Once deployed, the sensors must work unttended. Although, the sensors my hve fbriction-time identities, they should be treted s if they were nonymous s it is either imprcticl or infesible to devote ttention to individul sensors; The sensors re ssumed to be sttic (immobile) nd, t lest initilly, unwre of their loction; Ech sensor hs mximum trnsmission rnge, denoted by t x, ssumed to be much smller thn the width or the length of the deployment re. 1 This implies tht messges trnsmitted by sensor cn only rech recipients in its proximity; 1 Of course, t x depends on the prticulr type of sensors deployed. Under technology vilble t the time of this writing, t x is bout 30m for micro-sensors.

16 5 The reception circuitry of sensors is ble to determine the strength of received signls (RSS). This feture llows ech sensor to estimte its distnce to the trnsmitter. Unfortuntely, this estimte is often inccurte due to the irregulrity of signl propgtion resulting from surrounding noise, signl reflection nd refrction, nd multi-pth fding. This limittion is inherent to ll RSSIbsed distnce mesurements s mentioned in [8]. However, in our protocol, we try to get round this inccurcy by: () Mking distnce estimtes bsed on the verge vlue of the received signl strength of severl trnsmissions, (b) Avoiding ny RSSI-bsed distnce mesurements for long-rnge trnsmissions in order to reduce the probbility of being ffected by surrounding noise. II- Sinks: In ddition to the tiny sensors, the network contins smll number of sinks. These sinks re responsible for tsking the sensors nd for collecting the ggregted results; The sinks re much more powerful thn the tiny sensors. They hve no energy constrints either by being connected to stedy energy supply or by being powered by rechrgeble btteries; As they do not hve energy constrints, the sinks re ssumed to be wke ll the time; The sinks my be sttic or mobile, however they re lwys wre of their positions (i.e they might be equipped with GPS devices or their positions cn be entered mnully); Ech sink is equipped with two trnsceivers with trnsmission rnge T x tht exceeds the mximum trnsmission rnge t x of sensor. The first trnsceiver is omnidirectionl while the second is unidirectionl with nrrow bem width. The sink nodes re wre of their orienttion, nd cn rotte their unidirectionl trnsmitter to ny ngle [0, 2π] to trnsmit in ny direction.

17 6 I.4 ROADMAP The reminder of this disserttion is orgnized s follows, in Chpter II, we briefly summrize relted work pertining to reserch to solve chllenges in sensor networks. Then in Chpter III, we present the detils of our bckbone construction protocol. Strting from section III.5, more emphsis is given to show how the proposed bckbone cn help mitigte mny of the typicl chllenges in sensor networks including sensor locliztion, dt clustering, dt ggregtion, nd geogrphic routing. Chpter IV is dedicted for bckbone-bsed tsk ssignment nd workforce selection. We propose one centrlized nd nother distributed protocols tht cn be implemented on top of the proposed bckbone. The chpter is concluded by fmily of dt ggregtion protocols tht cn be integrted with the proposed tsk mngement process. In Chpter V, we present bckbone guided sleep scheduling scheme tht cn be used to blnce sensor energy consumption. Finlly, in Chpter VI, we conclude our work nd highlight on future reserch directions.

18 7 CHAPTER II STATE OF THE ART Although, dvnces in technology enble the mssive production of inexpensive sensors tht cn be deployed in lrge geogrphicl res, it rises numerous chllenges on the protocols needed to interct with these sensors efficiently. In this chpter, we provide n overview of the work tht hs been reported in the literture nd tht is relted to the construction of bckbones in sensor networks. We lso highlight on the stte of the rt techniques which were proposed to solve the most importnt chllenges in sensor networks including locliztion, scheduling nd routing. II.1 SENSOR BACKBONES AND INFRASTRUCTURES Wd et l. [9] proposed virtul infrstructure for mssively-deployed collection of nonymous sensor nodes. They defined coordinte system tht provides n interesting clustering scheme for nonymous sensors, nd referred to the process in which sensors lern their coordintes s sensor trining. Sensor trining cn be repeted in scheduled or d-hoc bsis to provide robustness nd dynmic reorgniztion. During the trining process, the deployment re round ech sink is divided using number of equingulr sectors (wedges) nd concentric circles (corons) centered t tht prticulr sink. Olriu nd Stojmenović [10] hve shown tht the rdii of the corons cn be determined to optimize the efficiency of sensors-to-sink trnsmission. The group of sensors which reside in the region determined by the intersection of specific coron nd specific wedge mps into one cluster. Although sensor trining cn simplify mny of network mngement tsks like routing nd dt fusion, it hs n inherent sclbility shortcoming. In prticulr, s we move from the sink node outwrd, the cluster sizes increse from one coron to the next. After certin point, sensors within the sme cluster my not be within the communiction rnge of ech other, hence more complex dt fusion nd leder election protocols re needed to hndle these clusters. More recently, Bertossi et l. [11] hve enhnced the trining protocol presented in [9]. The new pproch outperforms the originl pproch in terms of the overll time for trining by lowering it from liner to squre-root function of the size of the coordinte system used for loction wreness.

19 8 Srinivs et l. [12] presented novel hierrchicl wireless networking pproch in which some of the nodes re more cpble thn others. In their model, the more cpble nodes serve s mobile bckbone nodes nd provide bckbone over which end-to-end communiction cn tke plce. In their pproch they try to control the mobility of the bckbone nodes in order to mintin connectivity. They formulte the problem of minimizing the number of bckbone nodes nd refer to it s the Connected Disk Cover problem. They show tht it cn be decomposed into the geometric disk cover (GDC) problem nd the steiner tree problem with minimum number of Steiner points (STP-MSP). They provide pproximtion solutions to both problems. Although, mobile bckbones cn solve mny of network mngement problems, they ssume the existence of nodes with more dvnced cpbilities which my not be vilble in mny sensor network deployment scenrios. Frey et l. [13] showed tht ny sensor network grph cn be trnsformed into n ggregted form which is virtul overly grph (e.g. n infinite mesh of regulr hexgons) whose nodes re the centers of ll nonempty clusters (hexgons). Two nodes C 1 nd C 2 of tht overly grph re connected by n edge if there exists t lest one connected pir of network nodes with one node locted in C 1 nd the other locted in C 2. In their work, Frey et l. show tht network connectivity does not suffer from generlizing the concept of sensing coverge to rbitrry clusters. Hence, geogrphic routing cn be used on cluster-bse nd not on node-bse. II.2 SENSOR LOCALIZATION Most WSN pplictions require in wy or the other to ssocite sensor redings with the geogrphic loction in which the redings were tken. Getting loction informtion through recording positions mnully or through n expensive GPS chip re not vlid options for sensor networks. To ddress the locliztion problem nd to estimte good pproximtion of the position of ech sensor node, mny techniques hve been developed, ech of them hs its merits nd demerits. Up to the time of writing these lines, there is no specific lgorithm on top of others. Hence, we briefly summrize the technicl foundtions of the most importnt locliztion techniques proposed in the literture. In generl, locliztion techniques cn be clssified into two min ctegories: Rnge-bsed nd Rnge-free. Rnge-bsed techniques depend on rnge estimtion between nodes tht know their positions nd nodes tht do not. Rnges re

20 9 usully estimted bsed on mesurements of received signl strength (RSS), time of rrivl (TOA), time difference of rrivl (TDOA) nd ngle of rrivl (AOA)). Although Rnge-bsed techniques re more ccurte thn Rnge-free techniques, they shre common drwbck, they require dditionl hrdwre to be vilble with ech node to be ble to tke the required mesures. Rnge-free techniques hve been proposed to overcome the stringent hrdwre requirements of rnge-bsed techniques. The min ide behind these techniques is to exploit rdio connectivity informtion mong neighboring nodes to infer rough estimtes of their positions without tking ny rnge mesurements. This wy rnge-free techniques eliminte the need to equip ech node with ny specilized hrdwre llowing the mnufcturing cost of these nodes to be low. In the Centroid method [14], sensor node estimtes its position s the centroid of the polygon whose vertices re positioned t the nchors it could receive messges from. Anchors re loction wre nodes with trnsmission rnge tht is usully longer thn the trnsmission rnge of regulr nodes. If nchors were positioned uniformly, locliztion error cn be reduced however this cn not be gurnteed in d hoc or non sttic deployments. An obvious problem of this technique it loclizes ll the nodes tht receive messges from the sme subset of nchors (which is typiclly lrge number of nodes) to the sme position which increses verge locliztion error of nodes. An enhnced however rnge bsed vrition of this technique evlutes node position s the centroid of nchor positions weighted by the strength of signls received from these nchors. Another similr pproch for complex shpes ws proposed in [15]. The APIT method [16] divides the deployment re into tringulr sections using nchor nodes. Ech sensor pplies n pproximte PIT test to decide whether it is inside ech possible tringle or not. After tht it uses grid scn lgorithm to estimte the mximum likelihood re within which it resides. In fct, APIT could chieve good locliztion ccurcy, however it hs two mjor drwbcks. The grid scn lgorithm is very time nd resource consuming especilly if the grid size is smll. Also, APIT does not loclize nodes outside the convex hull of the nchor nodes ccurtely. The Ad-hoc Positioning System (APS) [17] ws proposed to llow non-gps enbled nodes to estimte their loctions in hop by hop fshion. Three different methods were investigted, nd the DV-Hop lgorithm ws the best to perform in

21 10 most cses. In the DV-hop lgorithm, ech nchor sends messge tht contins its loction informtion to ll the sensors round it. Sensors tht receive these messges forwrd them to their neighbors nd so the messges re flooded through the whole network. Within ech messge, there is field tht indictes number of hops this messge hs been forwrded. Bsiclly, this field is initilized to 1 t the nchor node, nd incremented t ech hop. This wy, sensors cn determine how fr they re (in terms of number of hops) from different nchors. The verge distnce per hop is clculted nd ech sensor cn estimte its distnce to different nchors. After tht, sensor node cn use the estimted distnces between itself nd three or more different nchors to loclize itself. Although DV-HOP is known to be one of the best known rnge free protocols, it hs severl drwbcks: first, the flooding nture of the protocol consumes much of sensor energy, second, it is not lwys esy nd sometimes infesible to use triltertion to estimte node position using pproximte distnces to nchors. The Amorphous [18] locliztion protocol depends on similr ide. The loction coordintes of the nchor re flooded throughout the network with the number of hops to the source nchor trcked in ech messge. This enbles ech node to mintin list of hop-count to ech nchor long with the loction of tht nchor. Ech node tht does not know its loction cn use this list to estimte its loction. Unfortuntely, Amorphous still hs the sme issues of DV-HOP. Cricket [19], n indoor loction support system proposed by MIT, llows nodes to lern their physicl loction by using listeners tht her nd nlyze informtion from nchors. Anchors concurrently use rdio nd ultrsonic signls to send their loction informtion. The listener inference modules on the node overcome multipth nd interference nd improve locliztion ccurcy. AHLoS [20] is similr to Cricket nd uses RF nd ultrsound for indoor locliztion. TDOA techniques like these techniques rely on extensive hrdwre tht might be expensive nd energy consuming, mking it less suitble for sensor networks. Another drwbck of TDOA techniques tht use ultrsound, they require dense deployment s ultrsound signls propgte for limited distnce only. The lighthouse system [21] uses rotting nchor tht produces prllel light bem of fixed width. A sensor node detects the light bem for period of time tht depends on the distnce between the nchor nd the sensor. If the rottion speed nd the width of the bem re known to the sensors, then ech sensor cn mesure the time it detects the light bem nd estimtes the distnce nd the ngle to the

22 11 nchor. Acoustic-bsed rnging techniques like BeepBeep [22] ws proposed to loclize sensors bsed on the two-wy time of flight of the beeps between two communicting devices. Other specil purpose locliztion techniques hve lso been proposed. Underwter 2D nd 3D locliztion ws proposed in [23], [24]. APL ws suggested in [25] to ddress locliztion in Rod Networks where most Rnge bsed locliztion techniques fil due to the sprse nture of deployment. APL uses binry vehicledetection timestmps to obtin distnce estimtes between ny pir of sensors on rodwys. II.3 SLEEPING SCHEDULE To sve their energy, sensors spend its whole life switching between two modes, sleeping mode (power consumption is minimum) nd wke up mode (power consumption is reltively high). Different deterministic nd probbilistic schemes cn be used to determine the schedule bsed on which sensors sleep nd wke up. Next, we summrize the technicl foundtions of the most importnt pproches we found in the literture. In [26], new scheduling protocol ws proposed to mximize network lifetime for rre event trget surveillnce systems. The protocol provides schedule tht gurntees tht ech point in the environment is sensed within some mximum intervl of time, clled the detection dely. However the protocol does not hndle QoS requirements tht might require ech point to be monitored by k sensors (k > 1) for more relible redings. Another sleep scheduling protocol ws proposed in [27], however it requires synchroniztion between nodes which is hrd to chieve in sensor networks. Anlyticl nlysis of the mximum chieved upper bound on network lifetime ws presented in [28]. The uthors of [29] proposed blnced-energy scheduling scheme for clustered sensor networks. Their scheme ims to evenly distribute the energy lod of the sensing nd communiction tsks mong ll the nodes in the cluster, thereby extending the network lifetime. However, their scheme does not provide ny gurntees on QoS requirements expressed in terms of the number of sensors prticipting in ech tsk. Another lgorithm ws proposed in [30] for lrge-scle wireless sensor networks. The lgorithm llows ech sensor to probbilisticlly schedule its own ctivity bsed

23 12 on its node degree to gurntee minimum level of connectivity. Unfortuntely, k-connectivity does not imply k-coverge, nd so their lgorithm does not provide ny gurntees tht k-coverge bsed QoS requirements re stisfied.

24 13 CHAPTER III BACKBONE CONSTRUCTION PROTOCOL The min gol of this chpter is to describe our pproch for constructing network bckbone tht cn be used to simplify mny of the problems inherent to sensor networks. In Section III.5 nd in the following chpters, we show how the proposed bckbone cn provide solutions to mny of the typicl chllenging problems in sensor networks including locliztion, clustering, dt ggregtion, routing, scheduling, workforce selection mong others. III.1 COMMUNICATION BACKBONES The concept of network bckbone cn be brodly generlized in sensor networks to include ny subnetwork of sensors. Referring to ny of these bckbones, it is strightforwrd to relize tht the network cn be esily clustered by ssociting ech non-bckbone sensor to its closest bckbone sensor. Figure 1, shows the different clusters constructed when using rndomly selected bckbone sensors. The reder should note tht the clusters re bounded by the Voronoi digrm whose vertices re the selected bckbone sensors. Bckbone Sensor Non-bckbone Sensor FIG. 1: Clusters constructed from rndomly selected bckbone sensors Due to the generlity of the definition, one would expect typicl sensor network to hve lrge number of bckbones, lthough most of these bckbones would hve no prcticl vlue. Apprently, in order to be prcticlly useful, network bckbone

25 14 hs to stisfy certin conditions. Our purpose is to highlight on the minimum conditions bckbone needs to stisfy in order to be prcticlly useful for communiction purposes. This would definitely give us more insight bout the best strtegy to follow in order to construct such bckbones. A network bckbone would be prcticlly useful for communiction purposes if it stisfied the following conditions: 1. The bckbone subnetwork is connected so it cn be used to forwrd messges between different prts of the network. Moreover, to reduce the number of hops, the distnce between ny two bckbone sensors should be s close s possible to sensor mximum trnsmission rnge, t x. 2. Bckbone sensors re well distributed cross the deployment re. Any nonbckbone sensor is within the trnsmission rnge of t lest one bckbone sensor. Furthermore, non-bckbone sensors should be evenly distributed cross different bckbone clusters. For uniform sensor deployments, this implies tht the res of bckbone clusters re equl. 3. The number of bckbone neighbors for ech bckbone sensor is mximized in order to mximize the number of different communiction pths between ny two nodes. c t x t x 60 t x b FIG. 2: Hexgonl clusters mximize the number of bckbone neighbors Assuming extremely dense network, where it is possible to find t lest one sensor s close s possible to ny point in the deployment re, we re interested to find the geometric shpe of bckbone clusters which will stisfy the conditions mentioned bove. We strt by plcing our first bckbone sensor round the center of the deployment re (see Figure 2). From condition (1), since the distnce between

26 15 ny two bckbone sensors is t x, the tringle bc must be equilterl. Moreover, to stisfy condition (3), we need to mximize the number of bckbone neighbors round sensor which cn be done by replicting the tringle bc s shown in Figure 2 confirming tht ech bckbone sensor cn hve up to six neighbor bckbone sensors. Furthermore, Figure 2 shows tht the bckbone cluster round sensor is in fct regulr hexgon with side length equls t x 3. III.2 THE BACKBONE CONSTRUCTION PROTOCOL Since, in our proposed protocol, the network bckbone is constructed strting from the sink nodes outwrds, we find it more pproprite to strt by showing how the protocol works round single sink node. It will then become cler tht ech sink performs the sme bckbone construction in disk of rdius R D round itself. Consider n rbitrry sink; we re interested in setting up bckbone in disk D of rdius R D, (R D T x ), round this prticulr sink. We note tht, s rule, the re covered by the disk is smll frction of the deployment re. Nonetheless, to simplify the presenttion, we shll refer to the sensor network built on the sensors in this disk s the network.

27 16 FIG. 3: Prtitioning the re round the sink into sectors We summrize the min steps of our construction protocol [31, 32] in the following phses: Tiling phse: in this phse the disk D round the sink is tiled using set of identicl regulr hexgons which mkes the re round the sink look like beehive (see Figure 3); Bckbone selection phse: in this phse the closest sensor to the center of ech hexgon is determined. These sensors re referred to s bckbone sensors; Non-bckbone sensor clssifiction phse: fter being selected, bckbone sensors nnounce themselves s well s the hexgon they represent to other nodes. Non-bckbone sensors use the received signl strength to determine the hexgon to which they belong. Tiling the disk D strts t the sink outwrds. The first hexgon is positioned such tht the center of the hexgon coincides with the sink. The side length of the tiling hexgons is tken to be tx 3, where t x is sensor mximum trnsmission rnge.

28 17 In prctice, we replce t x by ˆt x = (1 δ)t x, where δ 0.1. The reson behind this is to llow the selection of bckbone sensors which re very close to the trget hexgon center however they reside on the other side of the center nd my not be reched if we used t x. Referring to Figure 3, the tiling continues by plcing hexgons side by side in π six different directions (i.e., 3π, 5π, 7π, 9π 11π, nd ). We refer to these ngles s orienttion ngles. As shown in Figure 3, the re is divided into six sectors. In ech sector, the hexgons re stcked in rows. In the first row there is only one hexgon (column), in the second row there re two hexgons, in the third row there re three hexgons, nd so on. We propose ternry coordinte system to uniquely identify the vrious hexgons in the tiling bove. Specificlly, the hexgon in column c of row r in sector s is uniquely identified using the tuple s, r, c. It is worthwhile to mention tht lthough severl ddressing schemes for hexgonl networks [33, 34] hve been proposed, our coordinte system seems to be more pproprite for our construction protocol. The reminder of this section is divided into the following subsections: in Subsection III.2.1 we describe how the necessry ngles nd ssocited trigonometric functions re theoreticlly computed nd prcticlly mesured; in Subsection III.2.2 we specify the order in which bckbone sensors re selected; finlly, in Subsection III.2.3 we present the technicl detils of the bckbone selection process. III.2.1 Computing nd Mesuring Angles <s,r,c> γ Y <S,r,1> Sink α θ X FIG. 4: Estimtion of the position of hexgon centers

29 18 For given sector s, the tngent of the ngle θ s,r,c subtended by the positive x-xis nd the line connecting the sink to the center S s,r,c of the hexgon s, r, c is tn θ s,r,c = r 3 tn πs + (2c r 2) 3 r. (1) 3 + (r 2c + 2) tn πs 3 To justify (1), we begin by evluting the coordintes (x, y) of the center S s,r,c of the hexgon s, r, c, where the sink is ssumed to be locted t (0, 0). Referring to Figure 4, the distnce between S s,r,1 nd the sink is rt x, nd the distnce between S s,r,1 nd S s,r,c is (c 1)t x, where t x is the distnce between the centers of ny two djcent hexgons. Moreover, the ngle α between the line connecting S s,r,1 to the sink nd the positive x-xis is the orienttion ngle of sector s, nd the ngle γ cn be evluted geometriclly to be α + 2π. With this premble out of the wy, we cn 3 evlute x nd y s follows nd x = rt x cos α + (c 1)t x cos γ = (2s 1)π rt x cos + (c 1)t x cos 6 (2s + 3)π 6 (2) However, reclling tht sin y = rt x sin α + (c 1)t x sin γ = (2s 1)π rt x sin + (c 1)t x sin 6 sin cos (2s + 3)π 6 (2s 1)π 6 (2s 1)π 6 ( sπ = sin 3 π ) 6 = sin πs 3 cos π 6 3 = 2 (2s + 3)π. (3) 6 cos πs 3 sin π 6 πs sin 3 1 πs cos 2 3, ( sπ = cos 3 π ) 6 = cos πs 3 cos π 6 3 = 2 + sin πs 3 sin π 6 πs cos πs sin 2 3, ( πs = sin 3 + π ) 2 = sin πs 3 cos π πs + cos 2 3 sin π 2 = cos πs 3, nd

30 cos (2s + 3)π 6 ( sπ = cos 3 + π ) 2 we cn rewrite equtions (2) nd (3) s x = t [ x r 3 cos πs 2 3 y = t [ x r 3 sin πs 2 3 Using (4) nd (5), combined, we cn write confirming tht (1) holds. = cos πs 3 cos π πs sin 2 3 sin π 2 tn θ s,r,c = r 3 sin πs 3 r 3 cos πs 3 = r 3 tn πs 3 πs ] + (r 2c + 2) sin 3 + (2c 2 r) cos πs 3 + (2c 2 r) cos πs 3 πs + (r 2c + 2) sin 3 + (2c r 2) r 3 + (r 2c + 2) tn πs 3 = sin πs 3,, 19 (4) ]. (5) We wish to point out tht if for ll s, (1 s 6), the vlues sin( πs ), nd 3 cos( πs ) re tbulted, then ech sensor cn redily evlute the coordintes (x, y) 3 of the center of hexgon < s, r, c > s well s tn θ s,r,c without evluting ny trigonometric functions. This is very importnt s it reduces the energy expended by individul sensors. Prcticl mesuring of the ngle between the positive x-xis nd the line connecting the sink to ny sensor is more chllenging. As mentioned in the network model, we ssume tht the sink is cpble of both omnidirectionl nd directionl trnsmission. The directionl ntenn t the sink hs smll bem-width nd cn be rotted towrd ny direction. Recll tht ntenn physics sttes tht the trnsmission pttern of directionl ntenn consists of mjor lobe which is oriented in the direction of the trnsmission nd severl smller (bck nd side) lobes [35]. The received trnsmission power is mximum t the center of the mjor lobe nd reduces s we go fr from the center. For the purpose this work, we cn simplify the ntenn trnsmission pttern by representing it s nrrow sector with smll ngle tht is divided in hlf by the trnsmission direction bem (see Figure 5). Initilly, the sink uses its omnidirectionl ntenn to send sequence of WAKEUP messges to wke up sleeping sensors so they cn mesure their ngle to the sink. Obviously, the number of WAKEUP messges should be sufficiently lrge so tht ech sensor within the disk D will receive t lest one copy of the messge. Moreover

31 20 Mjor Lobe Z Trnsmission Direction Minor Lobes Y Side Lobe X Bck Lobe () Actul Pttern θ (b) Simplified Pttern FIG. 5: Rdition lobes nd bemwidths of directionl ntenn pttern to sve sensors energy, we suggest tht the number of remining WAKEUP messges be specified within ech messge, so when sensor receives WAKEUP messge in n erly stge nd relizes tht there will be more WAKEUP messges to follow, it cn sve energy by turning off its sensory nd reception circuitry, switching to the sleep mode fter djusting its internl timer to wke up on time. In ddition to wking up sensors, the lst WAKEUP messge should lso provide some level of synchroniztion mong sensors. Although this kind of synchroniztion my not be ccurte due to different trnsmission, propgtion nd processing delys t ech node, the chieved level of synchroniztion (within few milliseconds) is more thn sufficient for our purpose especilly in the existence of the lrge delys due to the mechnicl rottion of the directionl ntenn. After the lst WAKEUP messge, ech sensor turns on its reception circuitry nd wits. At the sme time, the sink uses its directionl ntenn to strt trnsmitting ngle estimtion messges strting from n initil ngle θ 0. After trnsmitting messge, the sink rottes its ntenn by smll ngle θ, then it trnsmits the next messge nd so on. Although ngle estimtion messges re very short, they convey useful pieces of informtion to sensors. The most importnt mong these pieces is the current ngle of trnsmission θ. When sensor receives recognizble ngle estimtion messge (i.e received signl strength p r is lrger thn some threshold vlue p th ), it stores the ngle θ long with the power of received signl p r. When the ntenn of the sink returns bck to the initil ngle θ 0, it cn either stop, or strt

32 21 nother cycle using different vlue for the rottion ngle θ in order to enhnce the ccurcy of the estimted ngles. Obviously, there is trde-off between the ccurcy of the estimted ngles nd the number of cycles needed which will definitely ffect the time nd energy consumption. After the lst ngle estimtion cycle, ech sensor estimtes its ngle s the verge of the received ngles weighted by their received signl strength p r. Mthemticlly, this cn be written s θ = n m=1 p rmθ m n m=1 p, (6) rm where n is the totl number of received messges, θ m is the ngle trnsmitted in messge m, nd p rm is the received signl strength of messge m. Recll tht the power of rdio signls decys proportionlly with the inverse of the trveled distnce rised to the pth loss exponent ( 2). Typiclly, reflected signls trvel distnce tht is longer thn the distnce trveled by direct LOS signls. Hence, if the initil trnsmission power p 0 remins the sme, then the received power of reflected signls should be smller thn the received power of direct LOS signls. Consequently, when ngle θ m is weighted by the received power p rm, we reduce the impct of reflected signls on the ccurcy of estimted ngle. Algorithm 1 Evlute sin θ nd cos θ Input: Angle θ Output: sin θ nd cos θ 1: θ 2 = θ θ ; 2: F [] := { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 } ; 0! 1! 2! 3! 4! 5! 6! 7! 8! 9! 3: sin := F[9] ; 4: cos := F[8] ; 5: for (i=7; i > 0 ;) do 6: sin := θ 2 sin + F[i ] ; 7: cos := θ 2 cos + F[i ] ; 8: end for 9: sin := sin θ ; 10: return sin, cos ; After evluting its ngle to the sink using eqution (6), ech sensor evlutes the sine nd the cosine of its ngle using the well known McLurin expnsion. By dding the first 5 terms of the McLurin expnsion of the sine nd the cosine, we obtin n ccurcy of up to 4 deciml plces which is more thn sufficient for our purpose. Moreover, we cn rewrite the pproximted expnsions of the sine nd the

33 22 cosine functions using Horner s rule [36] to reduce the number of multiplictions s follows sin θ θ θ3 3! + θ5 5! θ7 7! + θ9 (((( 9! θ 2 9! 1 ) θ ) θ 2 1 ) ) θ θ 7! 5! 3! cos θ 1 θ2 2! + θ4 4! θ6 6! + θ8 ((( 8! θ 2 8! 1 ) θ ! 4! ) θ 2 1 2! ) θ Furthermore, we cn combine the evlution of the two series nd reduce required number of multipliction to 10 opertions only s shown in Algorithm 1. By implementing Algorithm 1, nd through 10 multiplictions only, ech sensor cn evlute the sine nd the cosine of its ngle to the sink node. We drw the ttention of the reder tht using this evlution method is more pproprite nd consistent with the limited computtionl power vilble to these tiny devices. III.2.2 Order of Selection of Bckbone Sensors The selection process of bckbone sensors strts when the sink selects the six bckbone sensors in the first row. After tht, the process continues recursively where the sensors in ny row select sensors in the next row. This continues for sufficient number of rows necessry to cover the desired disk D. In prctice, we expect the mximum number of rows to be typiclly round 5 rows. Bckbone sensors cn be selected in mny wys nd in different order. However, to void redundnt selections, minimize collisions, nd sve sensors energy, we propose set of rules tht determine the order nd the selection responsibility of bckbone sensors, these rules re: 1. Only the sensors with odd column coordinte re llowed to select. (i.e. sensor S s,r,c is llowed to select c 1 (mod 2). 2. If [(c = 1) nd (r 0 (mod 2))] Then S s,r,1 selects S s,r+1,1, S s,r+1,2 nd S s 1,r+1,r+1. Else S s,r,2c 1 selects S s,r+1,2c 1 nd S s,r+1,2c.

34 23 3. Assuming, tht the selection of single bckbone sensor tkes one time epoch, then to reduce collisions nd interference between trnsmissions, selection in odd sectors (i.e. s = 1, 3, 5) occurs in different time epochs thn selection in even sectors (i.e. s = 2, 4, 6). Figure 3 shows selection order nd responsibilities for sector 5. Reltive to the bove rules for bckbone selection we note the following: Bckbone sensors with even column coordinte do not select other bckbone sensors; Selection in odd rows in ny sector requires 2 time epochs, however selection in even rows requires 3 time epochs; The totl number of time epochs needed to serch in ll the six sectors is 4 time epochs for odd rows nd 6 time epochs for even rows. III.2.3 The Detils of Bckbone Selection In Subsection III.2.2, we specified the order nd the selection responsibilities ccording to which bckbone sensors re selected. In this section, we give the technicl detils of the bckbone selection process. The protocol strts when the serching entity (i.e. the sink or some sensor) uses eqution (1) to compute tn θ s,r,c, the tngent of the ngle subtended by the positive x-xis nd the line connecting the sink to the center of the trget hexgon identified by the tuple s, r, c. After tht, it brodcsts messge to ll the sensors in its neighborhood sking for the closest sensor to the center of the trget hexgon to declre itself. Recll tht sensors t this point re wre of their ngle to the sink node θ s s described in Subsection III.2.1. The sensors tht receive the messge, check if the difference tn θ s tn θ s,r,c is within cceptble rnge (i.e less thn certin threshold). The sensors within the rnge use RSS to estimte their distnce to the serching sensor S. They lso use dditionl informtion within the messge trnsmitted by S to estimte e 2 s, the squre of their distnce to the center of the trget hexgon (clcultion of e 2 s is presented in detil lter). After tht, ech sensor initilizes countdown timer to Ke2 s units nd wits till its timer expires. When the t 2 x timer of ny of these sensors expires, the sensor relizes tht it is the closest sensor to the center of the trget hexgon. Consequently, it brodcsts messge to ll

35 24 <i,1,1> e s Sensor S t x d Y s Sink θ ϕ α i Sector Orienttion Angle X s FIG. 6: Cse 1: estimtion of e s 2 its neighbors declring itself s the bckbone sensor representing the new hexgon. The sensors tht receive the messge stop their timers nd use this messge long with messges they receive from other bckbone sensors to determine the hexgon to which they belong. Sensors which evlute e 2 s to be lrger thn some threshold vlue (i.e. e 2 th ) do not initilize their internl timers. This provides the stopping criterion upon which the boundries of deployment re is reched. Although collisions might well occur, they do not represent big problem s ties between colliding sensors cn lwys be broken by ny contention-bsed mechnism. One wy of doing this is to let colliding sensors wit rndom mount of time before trnsmitting gin. The first sensor to trnsmit is selected to be the bckbone sensor. As we mentioned erlier, during the selection of bckbone sensors, ech cndidte sensor needs to estimte e 2 s, the squre of the distnce between the sensor nd the center of the trget hexgon. To evlute this vlue, we distinguish between two cses. The first cse hndles the evlution for the six bckbone sensors round the sink (i.e. in row 1), while the second cse hndles the evlution for bckbone sensors in other rows. Cse 1: for ech sector i, the sink hs to select bckbone sensor s to represent the hexgon i, 1, 1. The selection error of sensor s is e s nd it represents the distnce between s nd the center of the hexgon i, 1, 1. Bsiclly, the criterion to select s is to keep e s minimum. Using Figure 6, it cn be redily verified tht the position

36 25 <s b,r b,c b > e b Sensor b Sensor b position When ψ < п/2 t x d <s,r,c > l d ψ' ψ e Sensor Z b Z 2 = X 2 +Y 2 Y Sink ϕ θ θ b b X FIG. 7: Cse 2: estimtion of e b 2 (X s, Y s ) of sensor s reltive to the sink nd e 2 s cn be evluted s X s = d cos θ (7) Y s = d sin θ (8) e 2 s = t 2 x + d 2 2t x d cos(ϕ θ) = t x 2 + d 2 2t x d (cos ϕ cos θ + sin ϕ sin θ) = t x 2 + d 2 2t xd (cos θ + tn ϕ sin θ) 1 + tn 2 ϕ (9) where d is the distnce between the sink nd the sensor nd is estimted using RSSI. Cse 2: s shown in Figure 7, we ssume the existence of bckbone sensor tht ws previously selected by the protocol to represent the hexgon s, r, c. The selection error of sensor is denoted by e nd represents the distnce between nd the center of the hexgon s, r, c. We recll tht ws selected such tht e 2 is minimum. Now it is sensor s turn to select nother bckbone sensor b to represent the hexgon s b, r b, c b. Agin, b should be selected such tht the selection error e 2 b is minimum (e b is the distnce between b nd the center of the hexgon s b, r b, c b ). Our gol is to provide n expression for e 2 b tht cn be evluted by ech sensor

37 26 independently. The evlution process strts t the serching sensor,, when it brodcsts messge sking for the closest sensor to the center of the trget hexgon to declre itself. Within this messge, sensor includes: 1. Z = X 2 + Y 2, the Eucliden distnce between the sink node nd sensor ; 2. sin θ nd cos θ, the sine nd the cosine of the ngle between the positive x-xis nd the line connecting the sink to sensor ; 3. In ddition to this, sensor evlutes nd sends l nd tn ϕ b, where l is the Eucliden distnce between the center of the trget hexgon nd the sink, while ϕ b is the tngent of the ngle subtended by the positive x-xis nd the line connecting the sink to the center of the trget hexgon. Clerly, l nd tn ϕ b cn be evluted using l = X 2 + Y 2 tn ϕ b = r b 3 tn πs b + (2c r 2) 3, r b 3 + (rb 2c b + 2) tn πs b 3 where X nd Y re given by the equtions (4), nd (5) respectively. After receiving the messge trnsmitted by sensor, ech sensor continues the evlution of e b on its own. Given tht Z, sin θ, cos θ, l nd tn ϕ b re known from the messge received from sensor ; d is estimted through the strength of the signl received from sensor ; sin θ b nd cos θ b, (the sine nd the cosine of the ngle subtended by the positive x-xis nd the line connecting the sensor b to the sink) were estimted t protocol initiliztion through WAKEUP messges,

38 27 generic sensor b cn estimte the distnce Z b between itself nd the sink by elementry trigonometry s follows: where Z d = sin(π (θ b θ + ψ)) sin(θ b θ ) sin(θ b θ + ψ) = Z sin(θ b θ ) d sin ψ cos ψ + tn(θ b θ ) 1 + 2A tn ψ + A 2 tn 2 ψ 1 + tn 2 ψ = Z d = Z2 d A tn ψ + A 2 tn 2 ψ = Z2 d 2 + Z2 d 2 tn2 ψ, (10) A = 1 tn(θ b θ ) = cos(θ b θ ) sin(θ b θ ) = cos θ b cos θ + sin θ b sin θ sin θ b cos θ cos θ b sin θ (11) Becuse in our protocol we lwys ssume tht sensors in ny row select sensors in the next row, the vlue of the ngle ψ is lrger thn π 2. For the cse, when tn θ b = tn θ, which implies for our scenrio tht θ b = θ. Hence, ψ = π, nd Z b = Z + d. For the generl cse, we solve the qudrtic eqution (10) for tn ψ, After bit of lgebr, (A 2 d 2 Z 2 ) tn 2 ψ + 2Ad 2 tn ψ + d 2 Z 2 = 0 2Ad 2 ± 4A 2 d 4 4(A 2 d 2 Z 2 )(d 2 Z 2 ) 2(A 2 d 2 Z 2 ) = tn ψ tn ψ = Ad2 + Z d2 + A 2 d 2 Z 2 Z 2 A 2 d 2 (12) Here, we chose the negtive root to gurntee tht tn ψ is negtive since A > Z d. Now, we cn evlute Z b s, Z b sin ψ = d sin(θ b θ ) Z b = d sin ψ sin θ b cos θ cos θ b sin θ = d tn ψ (sin θ b cos θ cos θ b sin θ ) 1 + tn 2 (ψ) (13)

39 28 Finlly, fter evluting Z b, ech sensor cn pply the trigonometric lw of cosines to evlute e 2 b s e 2 b = l 2 + Z 2 b 2lZ b cos(ϕ b θ b ) = l 2 + Zb 2 2lZ b (cos ϕ b cos θ b + sin ϕ b sin θ b ) = l 2 + Zb 2 2lZ b (cos θ b + sin θ b tn ϕ b ) 1 + tn 2 ϕ b (14) After evluting e 2 b, ech sensor initilizes its internl timer using Ke2 b s described t 2 x erlier. The winning sensor, i.e. the sensor whose timer expires first, declres itself s the selected bckbone sensor by brodcsting messge to other cndidte sensors. The selected bckbone sensor estimtes its position reltive to the sink node using X b = Z b cos θ b (15) Y b = Z b sin θ b. (16) III.3 BACKBONE SWITCHING One of the mjor dvntges of our proposed bckbone is tht it provides n implicit clustering mechnism where hexgons cn be viewed s clusters nd bckbone sensors re cluster heds. Although this cn simplify mny network mngement tsks including dt ggregtion, leder election nd routing, it usully results in uneven energy consumption mong sensors. In prticulr, it imposes higher tsking lod on bckbone sensors much more thn it does on other sensors which results in depleting their energy much fster. One wy to overcome this problem is through chnging the cluster hed periodiclly in order to distribute the dditionl lod on different sensors. In this section, we propose bckbone switching s solution to the energy blncing problem. The min ide behind bckbone switching is to construct disjoint bckbones nd to periodiclly switch between these bckbones to blnce their energy consumption. Initilly, using the pproch described in Section III.2, ech sink constructs its first bckbone. For lck of better term, we refer to this s the min bckbone. As described erlier, the min bckbone is constructed by selecting sensors in six different π directions (i.e., 3π, 5π, 7π, 9π, nd 11π ). Now, wht hppens if the sink node rotted its positive x direction by some ngle θ such tht 0 < θ < π? Theoreticlly, 3 nd s shown in Figure 8, the rottion should result in selecting completely different

40 29 Sink Rottion Angle θ θ FIG. 8: Blncing energy consumption using bckbone switching set of sensors (i.e n lterntive bckbone). While theoreticlly correct, this view my not be entirely correct from prcticl perspective. Recll tht bckbone sensors re selected to be the closest sensors to the hexgon centers they represent. So it is possible tht even fter rottion, the sme sensor my still be the closest sensor to the new hexgon center, hence it cn be selected to be prt of the lterntive bckbone. Fortuntely, our selection protocol only selects sensors tht nominte themselves to be prt of the network bckbone (by prticipting in the countdown process). Hence, if sensors which re lredy prt of nother bckbone do not nominte themselves, they will not be selected s prt of the new bckbone giving chnce to other sensors to join the new bckbone. This trick provides simple solution to gurntee tht lterntive bckbones re disjoint. After constructing the min bckbone, ech sink constructs set of lterntive bckbones using ppropritely selected ngles θ i. Ech bckbone is ssocited with n ID ssigned by its sink node. At ny point of time, only one bckbone should be ctive. It is the responsibility of the sink to periodiclly brodcst messges to chnge the current ctive bckbone giving chnce to sensors in other bckbones to sve their energy.

41 30 No bckbone sensors in the shdow of the void region Void Region FIG. 9: Unloclized sensors in the shdow of void regions III.4 RECOVERING FROM SENSOR VOIDS In certin pplictions the sensors re deployed in rough environments in which there exist some spots in the deployment re where sensors cn not be deployed. We refer to these spots s voids. Voids cn be creted nturlly by physicl obstcles (e.g. lkes, strems, lrge rocks, deep holes, steep slopes, etc.). Voids cn be creted s well when ll the sensors within certin spot expire due to energy depletion. The min gol of this section is to discuss how our proposed construction protocol would perform in the presence of such voids. Initilly, we point out to the problems tht might rise due to the existence of these voids. After tht, we show how the proposed protocol cn overcome these problems. Figure 9 shows how voids cn prevent the propgtion of the bckbone selection process. If no bckbone sensors cn be selected in the void region nd, consequently, the selection process stops nd no bckbone sensors re selected in the shdow of the void region. To get round this problem, we propose the Even Neighbor Replcement rule nd enhnce it lter by dding Bckwrd Selection.

42 31 FIG. 10: Illustrting even neighbor replcement Even Neighbor Replcements No bckbone sensors in the shdow of the void region Void Region FIG. 11: Recovering from voids using even neighbor replcement

43 32 III.4.1 Even-Neighbor Replcement the Detils Recll tht in our bsic bckbone selection rules, in ny row, only sensors with odd column coordinte re llowed to select bckbone sensors in the next row. Moreover, s shown in Figure 10-(), every even neighbor cn receive selection messges trnsmitted from its two immedite odd neighbors. If the initil selection rules filed to select n odd bckbone sensor due to the existence of void, then ll bckbone sensors belonging to the tree rooted t the missing sensor re pruned out. This kind of behvior blocks the propgtion of bckbone selection in the shdow re behind the void. The ide behind the even neighbor replcement rule is to llow the immedite even neighbor sensors to replce ny missing odd sensors in order to continue the selection chin. Figure 11, shows n exmple of how our construction protocol would work when pplying the even neighbor replcement rule. Although, the protocol cn recover from the void region, its recovery rte is reltively slow which leves lrge region of the deployment re uncovered by bckbone sensors. This motivtes for our next selection rule tht we dd to our protocol rules in order to expedite the rte by which voids re recovered. III.4.2 Bckwrd Selection the Detils The ide behind this rule is simple. If the selection of bckbone sensor ws initited by sensor other thn the one determined by the bsic rules (i.e. through even-neighbor replcement or nother bckwrd selection), then the selection responsibility is reversed nd the newly selected sensor crries the responsibility of selecting the odd sensor tht ws supposed to select it. Figure 12 shows how the rule is pplied. Initilly, sensor S 2i 1 trnsmits messge clling for the closest sensor to the center of hexgon H 3 to nnounce itself. As expected, sensor S 2i receives the messge trnsmitted by sensor S 2i 1, however, it does not receive similr messge from sensor S 2i+1. This motivtes sensor S 2i to pply the even neighbor replcement rule nd trnsmits messge clling for the closest sensor to the center of the hexgon (H 2 ) to nnounce itself. Sensor A 2 receives the messge nd nnounces itself to other sensors. Moreover, from the informtion embedded within the messge tht triggered its selection, sensor A 2 relizes tht the messge ws trnsmitted by sensor

44 33 Bckwrd Selection H 1 T+3 Even Neighbor Forwrd Replcement Selections H 2 H 3 H 4 A 2 A 1 A 3 A 4 T+4 T+5 H 5 S 2i+1 T+2 T+1 S 2i S 2i-1 T Void Region FIG. 12: Even-neighbor replcement with bckwrd selection S 2i+1 nd not sensor S 2i s determined by the bsic selection rules. This motivtes sensor A 2 to pply bckwrd selection rule nd trnsmits bckwrd messge clling for the closest sensor to the center of the hexgon (H 5 ) to nnounce itself. Sensor S 2i+1 receives the messge nd nnounces itself to other sensors. Following the bsic selection rules, sensor S 2i+1 trnsmits messge clling for he closest sensor to hexgon H 1 to nnounce itself. Sensor A 1 responds to this messge nnouncing itself to other sensors. After fulfilling its forwrd selection obligtions, nd becuse it ws not selected ccording to bsic forwrd selection rules, sensor S 2i+1 continues its bckwrd selection towrd the void region. Figure 13, illustrtes n exmple of how our proposed construction protocol works when pplying the even-neighbor replcement rule with bckwrd selection. Obviously, the protocol cn recover from void region efficiently. III.5 MITIGATING NETWORK CHALLENGES Sensor networks hve their own distinguishing chrcteristics tht set them prt from other types of networks. The d-hoc nture of deployment, loction unwreness, modest non-renewble energy budget, limited computing nd communiction cpbilities, long with the dynmiclly chnging topology induced by the sleepwke cycles re only few exmples of the typicl chllenges fced by WSN protocol

45 34 Bckwrd Selection Even Neighbor Replcement Void Region FIG. 13: Voids recovery using even neighbor replcement/bckwrd selection designers. Insted of solving ech of the forementioned problems individully, fcing the sme common chllenges with ech problem, we show how our proposed bckbone cn be very useful in collectively simplifying solutions for these problems. Our network bckbone provides some form of virtul infrstructure tht llows the sensors to cquire corse-grin loction wreness nd promotes dynmic clustering. Thus, on the one hnd, the infrstructure provides the sensors with necessry informtion tht enbles them to ssocite their sensory dt with the geogrphic loction in which the dt ws mesured nd, on the other hnd, it simplifies the tsk of clustering the sensors in support of vrious network tsks. Once such n infrstructure is in plce, entire protocol suites cn leverge the infrstructure, resulting in ese of progrmming nd energy svings. In prticulr, by tiling the re round sinks using identicl hexgons, the construction lgorithm clusters sensors bsed on their loctions into hexgons (clusters). Bckbone sensors represent cluster heds nd cn ply crucil rule in dt ggregtion, workforce selection, tsk mngement, leder election, duty cycle scheduling, nd locl synchroniztion. In the following subsections, we show how our proposed bckbone cn simplify sensor locliztion [37], locl dt ggregtion, geogrphic routing, nd clustering. We lso point to how using mobile sinks on top of our bckbone cn reduce the energy holes growth rte within the network. We dedicte Chpter IV for bckbone-bsed tsk mngement nd workforce selection. We strt Chpter V by rigorous nlysis

46 35 on wke sensor density nd its reltion to different scheduling schemes. After tht, we propose bckbone guided energy-wre scheduling scheme for blncing sensor energy consumption. III.5.1 Sensor Locliztion As the exct position of sink (X sink, Y sink ) cn be brodcst to ll the sensors in disk D of interest, s n dditionl field in WAKEUP messges, we ssume without loss of generlity tht (X sink, Y sink ) is known to ll the sensors in D. Moreover, we hve shown erlier in Subsection III.2.3 tht the sensors cn estimte their position reltive to the sink through equtions (7),(8), (15), nd (16). Using the vilble informtion, ech sensor cn estimte its bsolute position (X,Y) s: X = Y = { Xsink + d cos θ for cse 1 X sink + Z b cos θ b for cse 2 { Ysink + d sin θ for cse 1 Y sink + Z b sin θ b for cse 2 As expected nd confirmed by simultion in Section III.6, the locliztion ccurcy decreses lmost linerly with the distnce between sensor nd the sink. When the sinks re mobile or in the cse of reltively lrge number of sinks, we cn use these sinks to enhnce the chieved level of ccurcy s follows: the sensors tht reside close to one of the network sinks will typiclly belong to hexgon tht hs smll row coordinte, hence these sensors should be loclized ccurtely. On the other hnd, sensors tht reside fr from the network sinks nd close to the boundries of the locliztion regions of different sinks will typiclly receive locliztion messges triggered by ech of these sinks. From the received messges, boundry sensors cn loclize themselves reltive to ech of these sinks. These sensors should estimte their finl position s the weighted verge of the positions estimted from ech sink individully. The weight of ech position is evluted bsed on the row coordinte of the hexgon tht contins the sensor reltive to the sink used to evlute this position. The positions ssocited with smll row coordintes re expected to be more ccurte, hence they re ssigned higher weights thn positions ssocited

47 36 with lrge row coordintes. Mthemticlly, this cn be expressed s follows, w i = 1 r i R + 1 X = S i=1 w i X i S i=1 w i Y = S i=1 w i Y i S i=1 w i where S is the number of sink nodes used to loclize the sensor, (X i, Y i ) is the position of the sensor s estimted through sink i, R is the mximum number of rows llowed within the disk centered t the sink, r i is the row coordinte of the hexgon tht contins the sensor when loclized through the sink i. III.5.2 Clustering nd Leder Election Our bckbone implicitly clusters the sensors bsed on their geogrphic loction. Ech hexgon represents cluster nd the bckbone sensor round the center of ech hexgon is the cluster hed which cn be lwys elected s the leder to coordinte between sensors in its hexgon for ny centrlized protocol. For instnce, bckbone sensors cn ply n importnt rule in workforce selection nd tsk mngement for ll sensing tsks issued in the hexgons they represent. More detils bout this pproch re presented in Chpter IV. Furthermore, bckbone sensors cn be treted s elected coordintors for ny centrlized synchroniztion or scheduling protocols for sensors within their hexgons. This is discussed in more detils in Chpter V. We note here tht when it comes to selecting the bckbone sensor in given hexgon, nothing prevents us from extending the selection protocol in the obvious wy to select committee of severl possible bckbone sensors tht my collectively ct s cluster leders or, indeed, my tke turns serving ny ssigned tsks. III.5.3 Geogrphic Routing Given the coordintes of the hexgon tht contins the source sensor in the ternry system defined by our bckbone protocol, it is strightforwrd to find pth from

48 37 the source hexgon to the sink by hopping through bckbone sensors representing the hexgons in between (see Figure 14). The detils behind the selection of the route through which dt flows towrd the sink re presented in the next subsection. Here, it is worthwhile to mention tht by controlling the mobility of sink nodes, we cn tremendously reduce the growth rte of energy holes within the network. III.5.4 Dt Aggregtion: Non-bckbone sensors within ny hexgon cn report their sensory dt to the bckbone sensor in their hexgon which cn loclly ggregte the dt before forwrding the ggregted result to the next bckbone towrd the sink node. Using bckbone hexgons, sensory dt ggregtion nd routing ggregted results towrd sink nodes cn be strightforwrd. In prticulr, when sensor prticiptes in ny tsk, it trnsmits its results to the nerest bckbone sensor where sensory dt cn be loclly ggregted. After tht, bckbone sensors in row r forwrd their ggregted results to bckbone sensors in row r 1 nd so on towrd the sink node. Bckbone sensors in row r use reversed version of the sme rules they followed during bckbone selection in order to determine which bckbone sensor in row r 1 to forwrd their dt to. Figure 14 illustrtes n exmple of dt ggregtion nd routing towrd the sink node. The reder should note tht our bckbone does not impose ny restrictions on the order or the type of locl dt ggregtion inside the hexgons. The ggregtion process runs completely under the supervision of the bckbone sensor within the hexgon. Furthermore, routing ggregted results towrd the sink node by reversely following the pth determined during bckbone selection utomticlly provides workround for routing problems tht might rise due to the existence of energy holes or void regions. III.6 SIMULATION RESULTS In order to evlute the performnce of our proposed bckbone, we hve built simultor tht implements our bckbone construction protocol. In our simultion, we hve run severl experiments using different network prmeters nd configurtions. In generl, we ssumed rectngulr deployment re where number of sinks were

49 38 Bckbone Sensor Locl Aggregtion Regulr Sensor Sink Inter-bckbone sensor ggregtion (Reversed selection rules) FIG. 14: Dt ggregtion nd routing through our bckbone plced uniformly cross the deployment re. We used stndrd uniform pseudorndom genertor to distribute sensors with required density in the deployment re. We estimted the bckbone selection error s the verge Eucliden distnce between the position of the selected sensor nd the position of the center of the hexgon it represents. Mthemticlly, Error = 1 n (xsi x hi ) n 2 + (y si y hi ) 2, i=1 where (x si, y si ) is the position of bckbone sensor representing hexgon i =, s, r, c nd (x hi, y hi ) is the position of center of the hexgon i. To verify the correctness of our simultion implementtion, we initilly run our simultion ssuming exct distnce nd ngle mesurements. Bsiclly, in the bsence of distnce nd ngle mesurement errors, there should be no errors in sensor locliztion (except for minor trunction errors), lso the closest sensor to the center of ech hexgon should be lwys selected to be the bckbone representtive of this hexgon (we intentionlly put sensors t these positions nd verified tht they re ppropritely selected). Figure 15 shows plot of verge locliztion error for different rows nd verifies the correctness of our implementtion.

50 E-07 Averge Locliztion Error (meters) 1.20E-07 Density 0.30 Density E-07 Density 0.20 Density E-08 Density E E E E Row Number FIG. 15: Averge locliztion error of bckbone sensors vs. row number for different network densities using single sink in the bsence of mesurements errors After verifying the correctness of our implementtion, we conducted severl experiments to test the performnce of our proposed bckbone in the existence of errors in distnce estimtions nd ngle mesurements. To ccount for errors in distnce mesurements due to the irregulrity of signl propgtion, we represented estimted distnce between the trnsmitter nd the receiver s rndom vrible tht follows the Gussin distribution with men equls the exct distnce nd stndrd devition equls 0.1 of the mximum trnsmission rnge (round 3m when t x = 30m). In similr fshion, we represented mesured ngle between sensor nd the sink node s Gussin rndom vrible with men equls the exct ngle nd stndrd devition equls 3 rdin degrees. In our first experiment, we were interested to know wht the ctul hexgons produced by our protocol look like. Figure 16 shows the ctul hexgons produced by our simultion when the deployment re ws set to (200m 200m) squre, single sink node is plced t (0, 0), network density ws set to ρ = 0.3 sensors/m 2, nd sensor mximum trnsmission rnge t x = 30m. Although the boundries of the hexgons re completely distorted s they do not look like hexgons, we re still ble to distinguish the spots they occupy. For the sme experiment, Figure 17 shows the positions of hexgon centers nd the positions of the corresponding bckbone sensors representing them.

51 40 FIG. 16: Actul hexgons produced by simultion After tht, we conducted severl experiments to mesure the impct of distnce/ngle mesurement errors on the locliztion ccurcy. Figure 18-() shows the verge locliztion error under different network densities (0.05, 0.10, 0.15, nd 0.20) ssuming only distnce mesurement errors. The figure shows tht the bckbone selection error increses lmost linerly with the row number. When we repeted the sme experiment under the sme simultion prmeters in the existence of ngle mesurement errors. Figure 18-(b) shows tht the verge locliztion error still increses linerly with the row number, however the slope of the curve nerly doubles its vlue in the distnce-noise cse. This cn be explined by the fct tht n ngle mesurement error of θ for sensor tht is wy from the sink node by distnce d, will result in selection error proportionl to d θ. So for the sme ngle error, locliztion error increses s the vlue of d increses (i.e the row number increses). We repeted the sme experiment third time, however this time we considered both distnce nd ngle mesurement errors. As expected, Figure 18-(c) shows similr liner reltionship but with lrger slope. We lso conducted nother experiment to compre the locliztion ccurcy