ASFALT: Ā S imple F āult-tolerant Signature-based L ocalization T echnique for Emergency Sensor Networks

Transcription

1 ASFALT: Ā S mple F āult-tolerant Sgnature-based L ocalzaton T echnque for Emergency Sensor Networks Murtuza Jadlwala, Shambhu Upadhyaya and Mank Taneja State Unversty of New York at Buffalo Department of Computer Scence and Engneerng 201 Bell Hall, Buffalo, NY 14260, USA. {msj3, shambhu, mtaneja}@cse.buffalo.edu Abstract We consder the problem of robust node deployment and fault-tolerant localzaton n wreless sensor networks for emergency and frst response applcatons. Sgnature-based localzaton algorthms are a popular choce for use n such applcatons due to the non-unform nature of the sensor node deployment. But, random destructon/dsablement of sensor nodes n such networks adversely affects the deployment strategy as well as the accuracy of the correspondng sgnature-based localzaton algorthm. In ths paper, we frst model the phenomenon of sensor node destructon as a non-homogeneous Posson process and derve a robust and effcent strategy for sensor node deployment based on ths model. Next, we outlne a protocol, called Group Selecton Protocol, that complements current sgnature-based algorthms by reducng localzaton errors even when some nodes n a group are destroyed. Fnally, we propose a novel yet smple localzaton technque, ASFALT, that mproves the effcency of the localzaton process by combnng the smplcty of range-based schemes wth the robustness of sgnature-based ones. Smulaton experments are conducted to verfy the performance of the proposed algorthms. 1 Introducton Wreless sensor networks (WSN) are beng wdely used n emergency montorng and frst response applcatons lke natural calamtes (storm, hurrcanes), forest fres, terrorst attacks, etc. [14, 21, 18, 11]. Such networks are often referred to as Emergency Sensor Networks (ESN) [9]. Localzaton s the problem of determnng the poston of each sensor node (mote) after beng deployed at an area of nterest. Localzaton s extremely mportant n WSNs as the nformaton collected by the sensor nodes s of very lttle use unless t s assocated wth the locaton of occurrence. Dstrbuted localzaton protocols for WSNs can be dvded nto two broad categores namely Beaconbased methods and Sgnature-based methods. Beaconbased methods [20, 1, 15, 3, 7] requre a few specal nodes called beacon nodes, whch already know ther absolute locatons va GPS or manual confguraton and are ftted wth hgh power transmtters. Remanng nodes estmate ther locaton by frst computng dstance/angle estmates to the beacon nodes, and then applyng trangulaton or multlateraton to these dstance estmates. Sgnature-based or beaconless schemes [5, 6, 2, 10], on the other hand, assume that nodes are dstrbuted n a non-unform fashon over the deployment area, and use ths non-unform dstrbuton as a sgnature to compute locaton by observng node neghborhoods. In ths paper, we study the problem of localzaton from the pont of vew of ESNs. Sensor node deployment n emergency applcatons s hghly localzed for each pont (over the emergency area) and the sze of the node group at each pont depends on the ntensty of the montored event at that pont. Due to such a non-unformty n node deployment, sgnature-based schemes are deal for localzaton n ESNs. Moreover, such schemes elmnate the need for costly beacon nodes and GPS devces and thus the sngle pont of falure problem. But, one problem wth sgnaturebased schemes s that they assume a fxed node dstrbuton over the deployment area (throughout the perod of the applcaton) and thus ther accuracy s affected by factors that change the exstng node dstrbuton. Nodes over a deployment area can be arbtrarly destroyed, dsabled or dsplaced, thus changng the prevously fxed node dstrbuton. Sgnature-based schemes have to take such dstrbuton changes nto account before localzaton, else they wll produce naccurate results. Here, we attempt to construct sgnature-based localzaton schemes that are robust aganst random node destruc-

2 ton/dsablement. We focus on two man factors, namely, 1) the ntal node dstrbuton over the deployment area, and 2) random node dsablement. To provde an effcent dstrbuton of sensor nodes durng an emergency, we need a wellplanned deployment strategy that s not only robust aganst the vagares of the emergency stuaton but also helps sgnature-based localzaton n a postve way. To acheve ths, we outlne an emergency level-based deployment strategy that effcently dstrbutes the sensor nodes over the emergency area by dvdng the area nto varous emergency levels dependng on the severty of the emergency at a pont. The process of node destructon durng an emergency can be modeled as a non-homogeneous Posson process, and the deployment strategy employs ths model to make deployment decsons. Next, to mprove the fault-tolerance of exstng sgnature-based localzaton approaches, we propose an mprovement n the form of a Group Selecton Protocol (GSP). Accordng to ths protocol, only healthy or vable groups of nodes are chosen for partcpaton n the localzaton process. Although GSP provdes mprovement n accuracy, t does not smplfy the complex localzaton mechansm of sgnature-based schemes. To overcome ths, we ntroduce ASFALT, a smple, fault-tolerant localzaton scheme that combnes the salent features of both beaconbased and sgnature-based scheme. ASFALT uses dstance measurements to groups of nodes n ts neghborhood and a smple averagng argument to compute locaton. Usng expermental results, we show that the performance and localzaton accuracy of ASFALT are better than that of standard sgnature-based algorthms, e.g., [6], especally n stuatons of arbtrary dsablement/destructon of nodes. The rest of the paper s organzed as follows: the next secton presents the case study of a sgnature-based localzaton technque. Secton 3 presents the emergency levelbased deployment strategy for ESNs and the Group Selecton Protocol (GSP). Secton 4 descrbes ASFALT: our fault-tolerant localzaton technque. Secton 5 presents the evaluaton results and n Secton 6 we revew some earler research efforts n ths drecton. Fnally, we conclude and present some drectons for future research n Secton 7. 2 Case Study: A Sgnature-based (Beaconless) Scheme for Localzaton In ths secton, we present the case study of a sgnaturebased (beaconless) localzaton technque proposed by Fang et al. Interested readers may refer to the complete artcle [6] for detals. 2.1 Deployment Model and Localzaton Ths localzaton technque employs a group-based deployment strategy n whch the entre deployment area s frst dvded nto a grd of n ponts. Then, nodes are deployed n groups of equal szes at each pont on the grd. The fnal poston of each node after deployment s assumed to follow some non-unform dstrbuton, e.g., Normal (Gaussan), wth mean as the pont of deployment. Thus, the average deployment dstrbuton of any mote over the entre regon, f there are n groups, s: f overall (x, y) = 1 n n =1 1 2πσ 2 e [(x x)2 +(y y ) 2 ]/2σ 2 The eventual goal s to get dstance estmates from the target node at locaton θ(x, y) to each of the fxed pont on the grd where nodes are deployed, so that θ(x, y) can be determned by multlateraton. Let a = (a 1,...,a n ) be a vector representng the neghborhood observaton of the target node,.e., a number of nodes from group G are n the neghborhood of the target node. Gven the number m of nodes deployed n each group G and the probablty dstrbuton functon (p.d.f) of the deployment, the probablty that a s observed by the target node at θ (where X s a random varable representng the number of nodes from G that are neghbors to the target node and all X s are mutually ndependent) s, f n (a θ) =Pr(X 1 = a 1 θ)...pr(x n = a n θ) Let, g (θ) be the probablty that a mote from group G can land wthn the neghborhood of the pont θ. Then, ( ) m f = Pr(X = a θ) = (g a (θ)) a (1 g (θ)) m a Let z represent the dstance from θ to the pont where group G s deployed. It s clear that g (z ) = g (θ). Usng a maxmum lkelhood analyss t can be shown that the above lkelhood functon, f, s maxmzed when, g (z )= a m Now, to compute the value of z from g (z ) (z = g 1 (g (z ))), we need a formulaton for g (z ). Fang et al. have used complex geometrc technques to formulate g (z ) (see [6]). As a result, g (z ), whch s an extremely complex functon, cannot be computed n an onlne fashon by the low power sensor nodes. To overcome ths problem, a table-lookup approach s used to fnd z gven a and m,.e., g (z ) s pre-calculated (sampled) n an offlne fashon for dscrete values of z, and stored n the form of a table n the mote s memory. Once a and m are known, a sensor node can fnd the most ( lkelhood ) value for z by lookng up the value of g (z ) = a m from the table. Dstances to at least three or more known ponts (z s) can then be used to compute θ(x, y) by atomc multlateraton.

3 G p (x,y ) G p (x,y ) G p (x,y ) z θ(x, y) Θ (x,y ) z z Θ(x, y) Target Node Sensor Nodes Group Head Target Node Sensor Nodes Group Head Target Node Sensor Nodes Group Head (a) (b) (c) Fgure 1. Effect of node destructon on the accuracy of sgnature-based localzaton approaches. (a) No nodes destroyed, Node n queston at θ(x, y) and G = m = a =15(b) No nodes destroyed, Node n queston at θ (x,y ) and G = m = a =8(c) 7 nodes destroyed, Node n queston at θ(x, y), G = m =15and a =8 2.2 Dsadvantages In ESNs, node dstrbuton can change due to factors lke node destructon/dsablement, faulty nodes, etc. contrary to the statc node dstrbuton assumpton n sgnature-based localzaton schemes. Fgure 1 shows how random node destructon affects localzaton n sgnature-based schemes. Fgure 1(a) s the base lne scenaro. In ths case, the dstance (z ) between θ and the pont of group deployment p can be computed correctly. But, the above sgnature-based method cannot dstngush between cases (b) and (c),.e., when a node at θ actually observes just 8 nodes from group G t wll compute the dstance between θ and p as z (as shown n Fgure 1(b)). But, t may be the case that t just hears from 8 nodes from group G because the remanng 7 nodes mght be dsabled and the correct dstance s stll z and not z (as shown n Fgure 1(c)). Table 1 shows an approxmaton of the functon g (z ), dscussed above, as a table of values. Assumng a group sze (m ) of 100, we can see from Table 1 that a dfference of even a sngle observed node can cause an error of roughly 12m n dstance estmaton to the correspondng deployment pont. To verfy the naccuraces ntroduced by such an approxmaton we conducted smulaton experments usng the J-Sm [16] smulaton envronment for wreless sensor networks. In ths experment, we smulate the sgnaturebased algorthm dscussed above and observe the effects of random node dsablement on the localzaton accuracy of the algorthm. The deployment area s a 600m 600m square grd consstng of 9 ponts, each havng 20 nodes dstrbuted around t. In each run of the smulaton, the fnal poston of each node s sampled from a two dmensonal Normal dstrbuton (µ = 0, σ = 50, R = 200m) and the transmsson range s fxed at 200m. In each run, k (k vares from 1 up to 15) nodes per group are destroyed n Table 1. Table of g (z ) values, R = 200, σ =50 z g (z )= a m every group and the locaton of every node n a partcular group s estmated usng the sgnature-based scheme dscussed above. The results of the experment are outlned n the plot n Fgure 2. Performance of the algorthm s measured as an average of the localzaton errors of all the nodes n that group. From the plot, we can observe that the average localzaton error ncreases as k ncreases. Another trend that we observe n ths plot s that at hgh values of k, the localzaton naccuracy ncreases less steadly. Ths shows that beyond a certan threshold, the dsablement of nodes has lttle effect on ncreasng the localzaton error. Moreover, the average localzaton error n the case of zero node destructon (.e., k =0) s just under 30m, whch s hgh. One reason for the low accuracy of ths algorthm, even when k =0, s because the complex contnuous functon g (z ) s approxmated by a table of dscrete values. Thus, to mprove the accuracy and effcency of sgnature-based schemes n emergency applcatons, we need to address two ssues: 1) mprove fault-tolerance aganst dsabled nodes and 2) reduce complexty. Snce the accuracy of sgnaturebased schemes depends on the ntal dstrbuton of nodes,

4 we frst need to formulate an effcent strategy for sensor node deployment n emergency applcatons. We address ths problem n the followng secton. Average Localzaton Error (meters) Number of Destroyed Nodes per Group Fgure 2. Plot of Localzaton vs. Number of Dsabled Nodes 3 Node Deployment n Emergency Stuatons Exstng scatterng-based (by arplane, fre truck, etc.) deployment strateges have several shortcomngs for use n ESNs. Frst, deployment areas under severe condtons have hgh probablty of node destructon as compared to areas under relatvely tranqul condtons. Thus, deployng equal szed groups or n one bg group unformly over the entre area wll not be very effcent n emergency stuatons. Ponts on the deployment area where the effect of the emergency s hgh requre more nodes as compared to areas where the effect of the emergency s less hostle. But, just randomly deployng hgh number of nodes at ponts wth greater emergences s also not a good dea because the network may end up losng more nodes and the applcaton may fal. Also, manual deployment s dffcult due to the hostlty, naccessblty and unpredctablty at the ste of the emergency. Another problem s that current localzaton schemes do not ncorporate any nodes or protocols to montor changes n node dstrbuton after deployment. Thus, a more rgd analyss s requred before deployng nodes over the emergency area. 3.1 Model for Node Destructon We model the phenomenon of node destructon at a pont durng an emergency as a stochastc tme process, whchs a process that can be descrbed by a probablty dstrbuton wth doman as the tme nterval of the process. In other words, t s a collecton of random varables ndexed by a set T (tme). Ths helps to quantfy the expected number of nodes that wll be destroyed at a pont durng the emergency and the ntal number of nodes that should be deployed at that pont as a result. Assume that the emergency area s dvded nto a rectangular grd. Each dot n the grd represents a deployment pont, say p. Defnton 3.1 A deployment pont s a pont on the terran where a node (or group of nodes) s planned to be deployed. The pont where a node actually resdes after deployment, not necessarly the same as the deployment pont, s called the resdent pont. Let (x,y ) be the coordnates of the pont p. Assumng that there are k deployment ponts p 1 (x 1,y 1 ), p 2 (x 2,y 2 ),..., p k (x k,y k ),wehavek groups of nodes, G 1,G 2,...,G k,whereg s to be deployed at p. Snce we are tryng to model the effect of external factors on node survval, we assume that sensor nodes can be dsabled only by external factors lke fre, temperature, force, etc. and not by nternal/self factors lke battery falures, component malfuncton, etc. Let, t a be the start tme of the applcaton and t b be the end tme of the applcaton. Thus, the entre perod of the applcaton s, t a,b = t b t a. Each deployment pont, p, s assocated wth an emergency level basedonthe severty of the emergency condton at that pont, as defned below. Defnton 3.2 An emergency level λ at any nstance for a deployment pont s defned as the average number of destroyed nodes n group G per unt tme at that nstance and the correspondng functon λ (t) :t N s called the generalzed emergency level functon. In the above defnton, a node s consdered destroyed or dsabled f t s not capable of communcatng wth any of the neghborng nodes. The probablty of the number of dsabled nodes n a group over a fxed perod of tme can be expressed as a Posson dstrbuton because these dsablements occur wth a known average rate (emergency level) durng that nterval and are ndependent of the tme snce the last node dsablement. Specfcally, the number of nodes dsabled n a group durng the perod of the applcaton can be modeled as a non-homogeneous Posson process. Thss because, the average rate of node dsablement (emergency level) may change over tme (between the start and the end of applcaton) as the effect of the emergency at that pont changes. Thus, the number of nodes dsabled n a group G deployed over a deployment pont p n the tme nterval (a, b], gven as N (b) N (a), s as shown n Eqn. (1), P [(N (b) N (a)) = k ] = f(k,λ a,b ) e λa,b = (λ a,b ) k k! where k =0, 1,... G and λ a,b s the overall emergency level for the deployment pont over the tme nterval (a, b]. (1)

5 As mentoned before, an emergency level at a pont cannot be assumed to be constant throughout the tme nterval (a, b]. Emergency level at a deployment pont ncreases over tme f the stuaton at that pont worsens or t can decrease as the stuaton subsdes. As a result, the overall emergency level λ a,b for the deployment pont can be defned n terms of the generalzed emergency level functon λ (t) as shown below, b λ a,b = λ (t)d(t). a 3.2 Emergency Level-based Deployment In ths secton, we descrbe a deployment strategy, called the emergency level-based strategy, that can be used to deploy sensor nodes n an emergency stuaton. The frst ssue that we need to address s how to assgn an emergency level to each deployment pont. An emergency scenaro s an accumulaton of varous events occurrng at varous ponts. Each deployment pont s assocated wth a sequence of events; each event produces a dfferent rate of destructon. For example, a forest fre emergency conssts of some areas that are drectly under a wall of fre where the destructon rate s the hghest. Some areas where the fre s out but are stll under the effect of burnng objects have lower rates of destructon. Whle others that are just under the nfluence of smoke mght have a much lower rate of destructon. The best way to determne emergency levels for the varous deployment ponts s by repeated controlled experments. Before actual deployment, an emergency can be carred out n a controlled envronment and the sequence of events at any deployment pont can be smulated for a fxed tme, say the tme of applcaton t a,b. A fxed large number of nodes, m max (explaned n Secton 3.2.1), are deployed ntally n groups for each pont and the number of destroyed nodes can be noted. Such experments can be repeated n tmes and the number of destroyed nodes (k j ) s measured n each run j. Gven a sample of n measured values of dsabled nodes (k 1,k2...kn ) for each deployment pont, wewsh to estmate the value of the emergency level λ a,b of pont. Usng a Maxmum Lkelhood Estmaton (MLE) analyss, one can derve the most lkely value of the emergency level for any deployment pont asshownneqn.(2). λ MLE = 1 n n k j (2) j=1 Next, we focus on the group sze or the total number of nodes to be deployed at each deployment pont Determnng Deployment Sze Defnton 3.3 The deployment sze m for any deployment pont assocated wth an emergency level λ a,b s the number of sensor nodes to be deployed at that pont. The deployment sze m for a deployment pont depends on the emergency level λ a,b at that pont and s determned as follows. The deployment sze conssts of two components. The frst, called the standard deployment (m s ), s a fxed applcaton specfc constant that s same for every group. The next component, called vared deployment (m v ), s determned by the rate of node destructon at the deployment pont and s proportonal to the overall emergency level at the pont,.e., m v λ a,b. Thus, the deployment sze m at a deployment pont s a combnaton of the standard deployment and the vared deployment components,.e., m = m s + mv. Intutvely, more number of sensors are requred at deployment ponts wth hgher emergency levels as compared to lower ones. Accordng to our quantfcaton of the deployment sze, as the vared component of the deployment sze s proportonal to the emergency level t wll make sure that areas wth hgher emergences receve a larger deployment sze. Moreover, the vared component m v of the deployment sze offsets the effects of node destructon at that pont. Let m max be an applcaton dependent upper bound on the maxmum number of nodes that can be deployed at any pont that depends on factors lke network densty, cost of nodes, prorty of coverage etc. Sensor nodes wll be deployed at each deployment pont n groups of sze equal to the deployment sze m f and only f m m max Herarchcal Deployment Every group G conssts of at least one node desgnated as the group head ether pror to deployment or postdeployment through votng-based technques. Group heads (or base statons) have been mportant components n the desgn of effcent montorng applcatons rght from the ncepton of wreless sensor networks. Due to the low computaton power and storage capacty of sensor nodes, sensor network applcatons normally employ a record and forward paradgm [19]. In ths paradgm, sensor nodes forward data to ther respectve group heads as soon as t develops, whch then aggregates t and forwards t up the herarchy. Because of such a herarchcal desgn, group heads are aware of all the actve nodes n the group. Such a herarchcal desgn can be used n sgnature-based localzaton schemes to decde whch groups have suffcent number of nodes to perform localzaton accurately. But, the group head n such a settng can also be a sngle pont of falure. To overcome ths problem, a group can appont more than one group head dependng on factors lke sze of group, dstance between deployment ponts, type of applcaton, etc. But, to elucdate the current exposton, we assume wthout loss of generalty that each group conssts of a sngle, always on (.e., t s never dsabled) group head.

6 We now summarze the deployment strategy: 1. Dvde the deployment area nto a fxed set of deployment ponts. 2. Assumng that there are k deployment ponts, assgn an emergency level to each deployment pont as dscussed before. Then, prepare k groups of nodes, each of sze determned by the correspondng emergency levels. 3. All of the above nformaton lke the group szes, emergency levels, node dstrbuton (dscussed later) etc., called predeployment nformaton, s loaded nto the memory of every node before deployment. 4. Fnally, deploy each group of nodes at the correspondng deployment pont usng non-manual technques lke aeral scatterng, dsperson from a fre truck, etc. 3.3 Deployment Dstrbuton For a group of nodes thrown at a deployment pont, the probablty that the fnal poston of a node from the group s at the deployment pont s the hghest and the probablty decreases as we move away from the deployment pont. As a result, the fnal poston (resdent pont) of the nodes after deployment can be modeled as a contnuous random varable wth a certan fxed non-unform p.d.f lke Normal (Gaussan) dstrbuton as shown n Eqn. (3). Moreover, random varates wth unknown dstrbutons are often assumed to be Normal (Gaussan), especally n physcs and natural scences, and thus we can assume that the node dstrbuton around a deployment pont s Normal. For a group G, the mean (µ) of the p.d.f s the correspondng deployment pont p (x,y ). The standard devaton (σ) s applcaton specfc and depends on the coverage requred around the deployment pont. f (x, y) = 1 2πσ e [(x x)2 +(y y ) 2 ]/2σ 2 (3) Equaton (3) gves the probablty that a node n the group G has a fnal poston (x, y). LetPr (v) be the probablty that a node v selected at random belongs to the group G. Then, m Pr (v) = (4) m 1 + m m k where m, =1...k s the deployment sze of the group G. Thus, the overall dstrbuton of a randomly selected node v,.e., the probablty that the node v s present at the pont (x, y) on the deployment area s: f overall (x, y) = k Pr (v) f (x, y) (5) =1 Equaton (5) represents the probablty dstrbuton of the fnal poston of nodes just at the moment they are deployed. In theory, the probablty that a randomly selected node les closer to deployment ponts wth hgher emergency levels s hgh. But n practce, ths may not be true as nodes n groups near hgher emergency levels may also be destroyed wth a hgher probablty and as a result the actual sze of such groups may be farly smaller than ther orgnal sze at deployment. As dscussed n Secton 2.2, any scheme that uses ths dstrbuton should account for the loss of nodes n each group and use the most current group sze. Next, we dscuss a very smple and ntutve soluton to the above problem, called the Group Selecton Protocol (GSP). GSP, whch s mplemented on top of a sgnature-based localzaton algorthm, montors changes n node dstrbuton over the deployment area and helps to mprove the accuracy of the resultng localzaton schemes. 3.4 Improvng Sgnature-based Localzaton: Group Selecton Protocol (GSP) Let a be the number of nodes from group G that the target node at pont θ(x, y) can hear from and let z be the dstance from the target node to the deployment pont of group G. The problem wth the localzaton algorthm dscussed n Secton 2 s that n ESNs, not every observaton a n {a 1,...,a n } s correct or accurate. Groups where the node destructon rate s hgh mght not be able to provde the correct value of a for localzaton. One way to overcome ths problem s by beng selectve n choosng groups G s (and the correspondng observatons a s) for the localzaton process. We use a s from only those groups that are healthy. Defnton 3.4 The health of a group s quantfed by the number of actve nodes n the group. A node s actve f t s able to communcate wth at least one other node n the same group. In other words, only observatons from those groups are used durng localzaton n whch the current health of the group s at least equal to the standard deployment sze (m s ). Ths modfcaton wll reduce the number of z s (dstances) avalable for localzaton. But, as long as we have at least 3 relatvely accurate values of z s, localzaton can be done effcently. Absence of at least 3 values for z wll cause localzaton to fal, but due to the crtcalty of the applcatons n emergency stuatons sometmes no locaton s better than an ncorrect value. In ths protocol, group heads are used to montor the health of ther correspondng groups. After deployment, as the ad hoc network comes up, nodes begn sendng ntal setup nformaton to ther respectve group heads. Usng these communcatons from members of the group, the group head updates the health of ts group.

7 At regular ntervals, the group head broadcasts the current health of ts group. These broadcasts are forwarded by all nodes up to a certan hop count so that even nodes farther away can know the health status of a partcular group. The communcaton between nodes and the group head health broadcasts can be synchronzed wth the sleep-wake cycles of the nodes to save power. The group selecton protocol s as outlned n Algorthm 1. 1: Observe the neghborhood,.e., {a 1,a 2...a k a s the number of nodes n group G that are n rado range. } 2: Wat and observe health broadcasts (h ) from the group heads. Update h to the latest value for each group. 3: for all groups G for whch h s known do 4: f The group s healthy, say (h m s ) then 5: Compute g(z )=a /h. 6: Compute z from g(z ) by table look-up. 7: end f 8: end for 9: f z correspondng to at least 3 dstnct groups G s known then 10: Compute θ(x, y) by multlateraton (usng z s and ther correspondng p s) 11: else 12: prnt Cannot do Localzaton! 13: end f Algorthm 1: Group Selecton Protocol (GSP) Although the GSP proposes only mnor and ntutve mprovements to the process of sgnature-based localzaton, t performs better than exstng algorthms n dynamc scenaros. We verfy ths clam usng smulaton experments as outlned n Secton 5. Smulaton results show that GSP does mprove the localzaton accuracy of sgnature-based algorthms when nodes over the deployment area are randomly dsabled. Despte ths mprovement, there are some glarng problems wth current sgnature-based approaches that are stll left unaddressed by just employng the GSP. Current sgnature-based schemes are extremely complex nvolvng hard to compute functons. Smplfyng the process by usng regresson-based or table-based approxmaton technques results n loss of accuracy n addton to ssues lke offlne computaton and storng the functon as a table n the memory. The GSP provdes some mprovement n terms of accuracy relatve to standard sgnature-based approaches, but does not mprove on the complexty of such schemes. Moreover, GSP does not work well f node destructon s not localzed to only some deployment ponts n the network. To overcome these problems, we propose a smple fault-tolerant sgnature-based localzaton approach called ASFALT. 4 ASFALT: A Smple Fault-tolerant Sgnature-based Localzaton Technque In ths approach, nstead of just observng ts neghborhood, the target node computes dstances to every node n ts neghborhood. The set of dstance estmates from the target node to all nodes n a partcular group s called the dstance vector for that group. Ths dstance vector s a sample from the two dmensonal Normal dstrbuton wth mean as the dstance between the target node and the deployment pont of the group. Thus, gven a dstance vector, the dstance from the target to a deployment pont can be easly estmated by computng the mean of the sample. 4.1 Assumptons We assume that nodes are deployed over the deployment area usng an emergency level-based deployment strategy (Secton 3). Also, any node s effcently able to estmate ts dstance to ts one hop neghbors usng technques lke Receved Sgnal Strength Indcator (RSSI), Tme of Arrval (ToA), Tme Dfference of Arrval (TDoA), etc. [8]. Snce currently we are not modelng any specfc emergency, t s reasonable to assume that nodes are destroyed n a random fashon wthn a group. Ths s dfferent from the number of nodes destroyed whch s stll a Posson process and depends on the rate of destructon at that pont. All the symbols and termnology used n ths secton are same as Secton Localzaton Scheme Let M be the target node for whch localzaton has to be done and let θ(x, y) be the actual poston of M. TheAS- FALT localzaton technque s outlned n Algorthm 2. Let z be the actual dstance between θ(x, y) and the deployment pont. Let d 1,d2...dm d j R be the dstances of the nodes from the deployment pont (d j > 0, fthe poston of node j s after on the real lne and d j < 0 otherwse). Assumng that all m (m s + mv ) nodes n G are n the rado range of M, letz 1,z2 be the dstances...zm of the nodes from M (dstance vector). As mentoned before, the dstances n the set {d 1,d2...dm } follow a Normal dstrbuton and let d be the random varable that takes values n ths dstrbuton. Thus, E( d) =0 (6) In other words, the mean of all dstances selected from ths dstrbuton s 0. Let Z be the random varable that takes values n the dstrbuton followed by the dstance estmates n the dstance vector for group G. Snce each z j depends

8 on the correspondng d j, from Eqn. (6) we can clam that, E( Z )=z (7) In order to compute θ(x, y), M needs dstances z s to 1: Observe the neghborhood,.e., {a 1,a 2...a k a s the number of nodes from group G n rado range. }. 2: for all groups G for whch a 0do 3: Compute z 1,z2...za. 4: Observe health broadcasts (h ) from the group head. Update h to the latest value for the group. 5: end for 6: for all groups G for whch h s known do 7: f The group s healthy, say (h m s ) then 8: f (a <α ) then 9: Contnue; {Suffcent samples not avalable for approxmatng z } 10: else f (a α ) and (a <β ) then 11: Compute z = max{z 1,z2...za }; {Samples for approxmatng z do not cover the entre dstrbuton} 12: else f (a β ) then 13: Compute z = 14: end f 15: else 16: f (a <β ) then 17: Contnue; 18: else a a j=1 zj a {Compute mean} j=1 19: Compute z = zj a 20: end f 21: end f 22: end for 23: f z correspondng to at least 3 dstnct groups G s known then 24: Compute θ(x, y) by multlateraton (usng z s and ther correspondng p s) 25: else 26: prnt Cannot do Localzaton! 27: end f Algorthm 2: ASFALT Localzaton Algorthm at least 3 or more deployment ponts so that multlateraton can be done correctly. M frst observes ts neghborhood (a 1,a 2...a k ). Then, M computes the k dstance vectors {(z1 1,z2 1...za1 1 ), (z1 2,z2 2...za2 2 )... (z1 k,z2 k...za k k )}. It then computes the correspondng z by takng the mean of the correspondng z 1,z2...za values,.e., z = a j=1 zj a (8) It s obvous that larger the sample sze a, better s the approxmaton for z. The best approxmaton s when dstances from all the nodes n a group are avalable. But, an entre dstance vector may not be avalable because of two reasons: 1) the whole group mght not be n rado range (Fgure 1(b)), or 2) some nodes n a group may be dsabled (Fgure 1(c)). Thus, we need to dstngush between these two cases and handle them separately. To do ths we mplement GSP on top of ths algorthm to montor group health. If the group s healthy (h m s ) but stll the target node hears from only a few nodes n a group; ths would mply that not all nodes n that group are n the rado range. Otherwse, f the group s not healthy (h <m s ), the usefulness of the observaton vector s determned by the number of nodes vsble (a ) and a parameter β dscussed next Determnng α and β The ASFALT algorthm dscussed above requres two parameters to determne f a dstance sample or vector (z 1,z2...za ) for any pont s large enough to approxmate the dstance z correctly. The mean threshold β for a group G s the mnmum number of dstance values requred n the dstance sample so that t correctly represents the orgnal dstrbuton of nodes around the deployment pont (Normal). If the sze of the observed sample s at least β then the algorthm computes the dstance z as the mean of the dstance values n the sample. If the sze of the observed sample s less than β then t means only part of the group can be heard by the target node and z s computed as the largest value of the dstances n the sample. Generally, β = ms 2 works well for most cases. The mnmum threshold α s the mnmum number of dstance values n the sample requred to make any reasonable estmaton of the dstance z. If the sze of the observed sample s less than α, we dscard that observaton from consderaton n the localzaton process. Ths prevents ncluson of erroneous measurements n the multlateraton process. α s generally assgned a low value. From our smulaton experence, we observe that α ms 3 works well for most cases. 5 Evaluaton In ths secton, we present a detaled evaluaton of the GSP and ASFALT mechansms usng the sensor network smulaton tool J-Sm [16] and compare ther performance to the sgnature-based scheme proposed by Fang et al. [6]. In these experments, deployment s done over a grd of 600m 600m, consstng of 9 deployment ponts 100m apart as shown n Fgure 3(a). Each deployment pont has around 20 nodes deployed around t, followng a 2D Normal dstrbuton wth mean (µ) as the correspondngdeployment pont and standard devaton (σ) as 50. Snce we just want to observe the effects of node destructon on the accuracy of localzaton algorthms, we assume that the deployment sze of every group s same,.e., m = m s =20. Each

9 (meters) (meters) 100, Normal Node Group Head 7 8 Average Localzaton Error (meters) Sgnature-based Sgnature-based wth GSP ASFALT Number of Destroyed Nodes per Group Average Localzaton Error (meters) Transmsson Range (meters) (a) (b) (c) Fgure 3. (a) Smulaton setup - topology and node deployment (b) Plot of Localzaton Error vs. Number of Destroyed Nodes (c) Plot of Average Estmaton Error vs. Transmsson Range group has a sngle group head. The estmaton error s the Eucldean dstance between the actual poston and the estmated poston of the node and s measured n meters. We plot the average of the estmaton error of all the nodes only from group 4. Ths s done to avod the boundary nodes, because localzaton errors n the boundary nodes are generally hgh due to lack of suffcent samples for localzaton. In the frst experment, we smulate the sgnature-based approach, sgnature-based approach wth GSP and the AS- FALT algorthm n dynamc envronments. In each smulaton run, k nodes are destroyed from groups 1 and 5 (marked wth dotted crcles n Fgure 3(a)), and the value of k vares from 1 up to 15. The smulaton setup s not statc,.e, the node postons are not fxed throughout the experment. Nodes n each group are reassgned new postons accordng to the 2D Normal dstrbuton at the start of each smulaton run. The transmsson range of each node s fxed at 200m. The mean threshold β s 10 and the mnmum threshold α s fxed at 5 for each group. We also assume that a group s healthy f ts advertsed health h dffers from the orgnal health m by at most 2. The smulaton results are depcted n Fgure 3(b). As we can see from Fgure 3(b), the ASFALT localzaton approach performs much better as compared to the other two approaches and the average localzaton error of ASFALT ncreases much less sharply as compared to the other two. As the number of dsabled nodes per group (for groups 1 and 5) ncreases the average localzaton error for all of the 3 algorthms ncreases. For lower number of destroyed nodes, the sgnature-based algorthm outperforms the GSP. Ths s obvous, as the GSP does not consder samples from groups 1 and 5 even when the number of destroyed nodes s low (but more than 2). The GSP performs margnally better than the sgnature-based algorthm when the number of dsabled nodes s hgh. We have also conducted smlar experments for σ = 100,.e., nodes are sparsely dstrbuted around the deployment pont. The trends n the performance of the algorthms s smlar to the one shown n Fgure 3(b), but the localzaton error s comparatvely hgher n ths case. In the second experment, we observe the effect of rado range on the performance of ASFALT. The results are as expected (see Fgure 3(c)). When the rado range ncreases, each node s able to cover a larger area and thus not only dstance samples of a larger sze are avalable from each group, but also more groups become avalable. As a result, the effect of node destructon s lesser when the node rado range s hgher. 6 Comparson wth Related Work Despte the advances n the area of localzaton technques for sensor network, the problem of fault-tolerant localzaton has not receved much attenton. Robust localzaton schemes n the presence of malcous nodes and erroneous range measurements exsts [12, 13]. But, the problem of localzaton n the presence of erroneous measurements s dfferent from the one n whch entre nodes can be dsabled after deployment. The most notable work n faulttolerant localzaton was by Tnós et al. [17]. They present a novel fault tolerant localzaton algorthm developed for a system of moble robots, called Mllbots, that measure the dstances between themselves and then use Maxmum Lkelhood Estmaton to determne ther locaton. In another related work, Dng et al. [4] propose a medan-based mechansm for reducng the effect of faulty sensor nodes n target detecton and localzaton algorthms. To the best of our knowledge, there has been no prevous work specfcally addressng effcent and fault-tolerant deployment strateges and sgnature-based localzaton schemes for ESNs.

10 7 Concluson and Future Work In ths paper, we have addressed the problem of faulttolerant node deployment and sgnature-based localzaton for ESNs. We have outlned an effcent strategy for node deployment n emergency applcatons, called the emergency level-based strategy. We have also proposed a smple enhancement to exstng sgnature-based approaches, called Group Selecton Protocol (GSP), that mproves localzaton accuracy by montorng changes n node dstrbuton. Fnally, we have proposed ASFALT, a novel yet smple, fault-tolerant technque for localzaton n ESNs that combnes the salent features of both tradtonal rangebased and sgnature-based approaches. Our smulaton results have shown that ASFALT performs better compared to other sgnature-based technques, especally n stuaton of hgh node destructon. The mprovement provded by GSP and ASFALT comes at the cost of extra communcaton (health status advertsement) and computaton (dstance estmaton) overhead. Further evaluaton s needed to compare the complexty and overhead of the proposed technques aganst exstng schemes and we ntend to complete ths as a part of future work. Moreover, n ths work we assumed an deal rado model (crcular coverage area) whch s not practcal. As a part of future work, we would lke to extend the current work to ncorporate more practcal rado propagaton models lke two-ray ground model, shadowng model, etc. References [1] P. Bahl and V. N. Padmanabhan. RADAR: an n-buldng RF-based user locaton and trackng system. In IEEE INFO- COM Conference Proceedngs, pages IEEE Communcatons Socety, March [2] J. Bruck, J. Gao, and A. A. Jang. Localzaton and routng n sensor networks by local angle nformaton. In MobHoc 05: Proceedngs of the 6th ACM nternatonal symposum on Moble ad hoc networkng and computng, pages , New York, NY, USA, ACM Press. [3] N. Bulusu, J. Hedemann, and D. Estrn. GPS-less low cost outdoor localzaton for very small devces. IEEE Personal Communcatons Magazne, pages 28 34, Oct [4] M. Dng, F. Lu, A. Thaeler, D. Chen, and X. Cheng. Fault-tolerant target localzaton n sensor networks. EURASIP Journal on Wreless Communcatons and Networkng, 2007:Artcle ID 96742, 9 pages, do: /2007/ [5] L. Doherty, L. E. Ghaou, and K. S. J. Pster. Convex poston estmaton n wreless sensor networks. In IEEE INFO- COM Conference Proceedngs, Anchorage, AK, Aprl IEEE Communcatons Socety. [6] L. Fang, W. Du, and P. Nng. A beacon-less locaton dscovery scheme for wreless sensor networks. In IEEE INFO- COM 05 Conference Proceedngs, pages , Mam, FL, March IEEE Communcatons Socety. [7] T.He,C.Huang,B.M.Blum,J.A.Stankovc,andT.F.Abdelzaher. Range-free localzaton schemes n large scale sensor networks. In The Nnth Annual Internatonal Conference on Moble Computng and Networkng(MOBICOM). ACM SIGMOBILE, August [8] J. Hghtower and G. Borrello. Locaton systems for ubqutous computng. Computer, 34(8):57 66, August [9] M. Jadlwala, S. Upadhyaya, H. R. Rao, and R. Sharman. Securty and dependablty ssues n locaton estmaton for emergency sensor networks. In The Fourth Workshop on e- Busness (WeB 2005), Venetan, Las Vegas, Nevada, USA, December [10] X. J and H. Zha. Sensor postonng n wreless ad-hoc sensor. networks usng multdmensonal scalng. In Proceedngs of IEEE INFOCOM 2004, March [11] C. A. R. Jr. Sensors bolster army prowess. SIGNAL Magazne, AFCEA s Internatonal Journal, [12] L. Lazos and R. Poovendran. Serloc: secure rangendependent localzaton for wreless sensor networks. In The 2004 ACM workshop on Wreless securty, pages 21 30, Phladelpha, PA, October ACM SIGMOBILE. [13] D. Lu, P. Nng, and W. Du. Attack-resstant locaton estmaton n sensor networks. In The Fourth Internatonal Symposum on Informaton Processng n Sensor Networks (IPSN 05), pages ACM SIGBED and IEEE Sgnal Processng Socety, Aprl [14] K. Lorncz, D. Malan, T. R. F. Fulford-Jones, A. Nawoj, A. Clavel, V. Shnayder, G. Manland, S. Moulton, and M. Welsh. Sensor networks for emergency response: Challenges and opportuntes. IEEE Pervasve Computng, Specal Issue on Pervasve Computng for Frst Response, Oct- Dec [15] N. Pryantha, A. Chakraborty, and H. Balakrshnan. The crcket locaton-support system. In The Sxth Annual Internatonal Conference on Moble Computng and Networkng(MOBICOM), pages ACM SIGMOBILE, August [16] A. Sobeh, W.-P. Chen, J. C. Hou, L.-C. Kung, N. L, H. Lm, H.-Y. Tyan, and H. Zhang. J-sm: a smulaton and emulaton envronment for wreless sensor networks. IEEE Wreless Communcatons, 13(4): , August [17] R. Tnos, L. Navarro-Serment, and C. Pareds. Fault tolerant localzaton for teams of dstrbuted robots. In In Proceedngs of IEEE Internatonal Conference on Intellgent Robots and Systems, pages , Mau, HI, [18] Y.-C. Tseng, M.-S. Pan, and Y.-Y. Tsa. Wreless sensor networks for emergency navgaton. Computer, 39(7):55 62, [19] M. A. Tubashat and S. Madra. Sensor networks: an overvew. IEEE Potentals, [20] R. Want, A. Hopper, V. Falcao, and J. Gbbons. The actve badge locaton system. ACM Transacton on Informaton Systems, pages , Jan [21] L. Yu, N. Wang, and X. Meng. Real-tme forest fre detecton wth wreless sensor networks. In Proceedngs Internatonal Conference on Wreless Communcatons, Networkng and Moble Computng, volume 2, pages , September 2005.