Вы находитесь на странице: 1из 9

High-density Wireless Sensor Networks: a new Clustering Approach for Prediction-based Monitoring

Pieter Beyens*+,Ann Nowe' and Kris Steenhaut*t * { DINF, ETR0)-CoMo Vrije Universiteit Brussel Pleinlaan 2, 1050 Brussels, Belgium iDepartment IWT Erasmushogeschool Brussel Nijverheidskaai 170, 1070 Brussels, Belgium Email: {Pieter.Beyens, Ann.Nowe, ksteenha}@vub.ac.be

Abstract- We propose a new cluster-based approach sensor node. Measurements (sometimes aggregated) are that simplifies prediction-based monitoring for homoge- forwarded by the sensor nodes to a specific collector neous, high-density wireless sensor networks composed of a node or sink via multi-hopping [4]. For the observer, the large number of small, power-restricted nodes. Prediction- sink is the access point to the sensor network. We assume based monitoring can increase the autonomous lifetime of a continuous data delivery model in which sensed data the network by reducing communication. In our clustering approach, the cluster-heads spatio- is delivered continuously at an observer-specified rate temporally correlate and predict the measurements of [ 2 ] . The observer can change the delivery rate and/or the cluster-members by executing their prediction model. the required precision by a sink-initiated broadcast. When the batteries of the sensor nodes are not reguRouting is on& done by the gateway nodes at the circumference of the clusters while the non-gateway nodes, which larly rechargeable or not replaceable (e.g. in a remote or are positioned between the cluster-heads and their gateway hostile environment), the system lifetime is limited and nodes, are allowed to turn off their radio communication power consumption is a major design factor [4]. Energy as long as their measurements satisfy the predictions efficient protocols are needed to increase the autonomous of their cluster-head. Timing off radio communication lifetime of a WSN i.e. to monitor an environment for a results in high energy savings and can greatly improve long time with the required precision and time resolution. system lifetime. Our main contribution is the description of this cluster- Since communication is expensive in terms of power ing approach while the prediction models are beyond the consumption, a straightforward way of saving energy is to turn off the radio communication of as many scope of this paper. sensor nodes as possible while still guaranteeing network I . INTRODUCTION connectivity and satisfying the observer's requirements. Large-scale wireless sensor networks (WSNs) offer To deal with this optimization problem, we propose a promising prospects for the application area of envi- new cluster-based approach that enables us to implement ronmental monitoring because they provide a multi- prediction-based monitoring. Dependent of the dynamics dimensional, spatio-temporal view of the phenomenon to of the environment, prediction can enabie us to tum the end-user or observer [ll (21 131. A wireless sensor off radio communication of a high number of nodes. nenvnrk (WSN) consists of a large number of battery- This will results in high energy savings and prolong the powered, small-sized devices, called sensor nodes, ran- autonomous lifetime of a WSN, CIustering simplifies the complex problem of predicdomly deployed inside the environmental phenomenon [4] [ 5 ] . Sensor nodes are capable of (locally) sensing the tion by creating sub-networks, or clusters, responsibIe phenomenon and reporting the measurements through a for their own local prediction (see Figure 1). The clusters wireless Iink to their neighbors i.e. the sensor nodes realize their preGiction by assigning leader nodes or which are within transmission range of the reporting cluster-heads, responsible for the execution of their local

0-7803-8801-1/05/$20.00 (~)2005 IEEE.

Cluster Logical structure for routing

0 Cluster-head 0 'way node 0 Non-gateway node

Pig. 1.

Clustering for prediction based monitoring

prediction model. The prediction model itself is built offline. Routing is the responsibility of the gateway nodes which are positioned at the circumference of their cluster. Non-gureway nodes that lie between the cluster-heads and their gateway nodes are allowed to stop sending their measurements and to turn off their radio communication as long as their measurements satisfy the predictions of their cluster-head. Therefore, when the correlation of intra-cluster data is high and the prediction model of the cluster-heads performs well, the total traffic will be highly reduced and can so improve the system lifetime. Via the sink, the observer receives the measurements of communicating nodes and can, by using the same prediction model that the duster-heads use, find an approximation of the sensed environment. We focus on the description of this new clustering approach while prediction models are beyond our scope. The outline of this paper is as follows: in Section I1 we give an overview of the related work, the clustering approach and algorithm is explained in Section 111, followed by the results (Section IV). The paper ends with the conclusion and future work (Section V). 11. RELATEDWORK

excessive network traffic and time delay. Furthermore, given the dynamic and error-prone nature of WSNs, centralized protocols are more likely to fail since some specific nodes, commands or data are usually of higher importance in a centralized protocol [ 6 ] . Examples of centralized clustering algorithms are [7] [8] [9J [IO]. Opposed to centralized clustering algorithms are strictly localized clustering algorithms, In a localized algorithm, as defined in [ 1 I], sensor nodes only communicate with a small set. of other sensor nodes e.g. their neighbors. Chan and Perrig [6] have narrowed this definition to stricfly localized algorithms to exclude algorithms that use an incremental "growing" techniques like the creation of a tree structure starting from one node (e.g. the' Banerjee and Khuller clustering algorithm [9] and the cluster tree architecture described in [12]). Strictly localized algorithms are particularly interesting since they are usually more scalable (sensor nodes only use local information) and robust (faulty local information has usually only a local impact). Strictly localized clustering algorithms can follow several heuristics to select their cluster-heads. Chatterjee et al. [ I31 have made the following classification: highestdegree heuristics, lowest-id heuristics and node-weight heuristics. With these heuristics, respectively, clusterheads are chosen according to their node degree [ 141 [IS] [ 161 [ 1 71, node id [ 1 S] [ 191 [20] and node weight' [21] [22] [I31 [23]. For simplicity, we will use the lowest-id heuristic for our algorithm, though, it can easily be changed into another heuristic. Apart from Chatterjee's classification, threshold-based cluster-head selection approaches have been proposed by Heinzelman et al. in LEACH [24] and by Chan et al. in ACE [6]. The ACE algorithm results in a formation with highly uniform, non-overlapping clusters. In the next subsection we review algorithms based on the graph theoretical concept of dominating set.

In this section, we start with an overview of existing E. Dominating set algorithms clustering algorithms. Secondly, we review dominating A WSN can be regarded as a graph G = (V:E ) where set algorithms with which clustering algorithms are V is the set of sensor nodes and E is the set of interstrongly related. Finally, we end this section with a short node connections. A dominating set in graph theory i s discussion on prediction-based monitoring. the set of vertices which, together with their neighbors, represent all vertices of the graph. For WSNs, we require A . Chsteririg algorifhms the dominating set to be connected to maintain network In a distributed environment, clustering algorithms can connectivity. Finding the minimum dominaiing set of a be divided in two major classes: centralized and slrictly unit-disk graph is known to be a NP-complete problem luculized algorithms. A centralized approach is not attractive for large-scale WSNs because the retrieval of 'The node weight is a weighted sum of several parameters (e.g. global network information to a central point may induce energy level and node degree)

189

[25]. Various algorithms e.g. [26] [27] [28] [29] [301 [31] have been proposed to tackle this problem. Recently, Carle and Simplot-Ryl [32] have used the dominant pruning algorithm of Dai and Wu [33] in WSNs to calculate the area-dominating set i.e. a set of nodes that fully covers the monitored area. Calculating the area-dominating set is only possible when nodes have a monitoring area i.e. an area in which a node can detect event e.g. the detection of an object or animal. Their algorithm is not applicable for environmental monitoring where sensor nodes sense the environmental conditions at single points. We therefore propose a novel clustering algorithm. Clustering is a way to find a dominating set of a WSN because the cluster-heads form a dominating set. A connected dominating set can be obtained by selecting gateway nodes which connect two cluster-heads with each other. In our algorithm, the relaying of packets is done by the gateway nodes and not by the cluster-heads. We therefore require that our gateways nodes form a connected graph. We now discuss prediction-based monitoring. In our algorithm, prediction-based monitoring will enable us to turn off radio communication of non-gateway sensor nodes.

0 Clusier-head 0 Gateway node

- Logical structure tor routing @ Sink

0 Non-gateway node

Fig. 2.

Overview

new strictly localized clustering algorithm provides the logical structure with which energy efficiency can be obtained by turning off the radio communication of as many nodes as possible. C. Prediction-based moni foring Before we proceed, we make our assumptions clear. We assume a large-scale, high-density and static WSN. Prediction-based monitoring is a way to reduce communication between nodes. When measureinents of The deployment of the small, power-restricted sensor nodes are spatio-temporally correlated, prediction can nodes in the environment is random. All sensor nodes be used to turn off the radio of nodes of which one have the same hardware but unique id's and have limited expects to predict their measurements. In [34], Goel et al. processing and communication capabilities. All sensor have applied the concepts of MPEG video compression nodes have the same fixed transmission range and know by regarding a WSN as a sequence of pictures where their fixed geographical location (preferably by a GPSpixels are the sensor node's measurements. Recently, free technique e.g. [37]). Sensed data is delivered conXu et al. have presented prediction mechanisms for tinuously to the sink by using multi-hopping. We assume object tracking WSNs [35] [36]. The implementation ofd packets are delivered in finite time by an efficient MAC prediction models for our algorithm is beyond the scope protocol. In this subsection we explain how routing is done of this paper. on top of our logical cluster infrastructure. In the next 111. CLUSTERING FOR PREDICTION-BASED subsection we introduce the clustering algorithm itself. MONITORlNG Figure 2 shows how the sensor network retrieves enIn this section we describe our new clustering ap- vironmental data @er clustering is done. In step (a), proach. Subsection HI-A gives an overview of prediction all nodes start sensing the environment. Non-gateway based monitoring in WSNs. Subsection-Ill-B explains in nodes and cluster-heads forward their sensed data to the gateway nodes. The gateway nodes route the sensed detail our clustering algorithm. data to the sink using multi-hopping. Cluster-heads reA. Over&w ceive the data of their cluster-members and execute Our goal is to increase the autonomous lifetime of their prediction model. Important for later, cluster-heads a WSN for environmental monitoring. We believe our also report their status as cluster-head to the observer,

190

0Cluster
for muthg

TABLE 1
0 Cluster-head
Gateway

- Logical structure 0

DIFFERENLE~ BETWEEN TRADITIONAL CLUSTERING A N D


CLUSTERING FOR PREDICTION-BASED MONITORING

0 Non-gateway node

i
1
Goal

1
1
heads

Traditional clustering
simplify routing

Clustering far predictionbased monitoring simplify prediction

no forwarding task
data fusion execution of prediction model fonvarding task

Fig. 3. Traditional clustering


nodes

T step (b), the cluster-heads send their predictions to n their cluster-members (dashed lines). Except from some periodical checks, non-gateway nodes are allowed to tum off their radio communication as long as their measurements correspond to the predictions of their cluster-head taking into account some observer-specific parameters like accuracy and error margin. Because the gateway nodes have a routing task, they never put their radio in sleep mode. Step (c) in Figure 2 is the view over time the observer gets via the sink. Initially (before step (b)), the observer receives measurements of a11 sensor nodes. After predictions are sent to the cluster-members by the cluster-heads (after step (b)), only incorrectly predicted measurements are delivered to the observer. By executing the prediction model on the cluster-heads using the received measurements, the observer can get an approximation of the sensed environment. The observer can adjust the accuracy and error margin by a sinkinitiated broadcast (d). At first, the broadcast message is forwarded by the gateway nodes to the cluster-heads, secondly, the cluster-heads send it to the non-gateway nodes when they have their radio communication on again (note that non-gateway nodes might periodically check for cluster-head messages). Regularly, the cluster infrastructure must be rebuilt in order to balance the tasks evenly among the sensor nodes. Our approach differs significantly from traditional clustering approaches in which clustering provides a logical structure for hierarchical routing (see Figure 3). Hierarchical routing simplifies routing in large networks by creating a hierarchy of clusters, responsible far their own routing. The clusters realize their routing by assigning cluster-heads, responsible for the maintenance of their cluster and the forwarding of inter-cluster data via gateway nodes. In addition, cluster-heads could perform

gateways

send data to cluster-head

could turn radio off

data fusion of intra-cluster data to decrease the number of inter-cluster data packets. In contrast with the clusterheads for hierarchical routing, our cluster-heads are relieved from their routing task. Table I clarifies fhe differences between traditional clustering approaches and our clustering approach. In the next subsection, we proceed with the detailed description of our clustering algorithm.

B. Clustering algorithm for prediction-bused monitoring


We divide the system lifetime in cycles. Each cycle the clustering algorithm is executed. In the following, we explain one execution cycle. All sensor nodes know their neighbors and their geographical coordinates. At the beginning of each cycle, sensor nodes are set to the free state and, for simplicity, are sequentially activated following the lowest-id heuristic e.i. the node with the lowest id will be activated first. After this activation phase, each node stays a free node or has been selected as a cluster-head, gateway node or non-gateway node. Note that other activation heuristics are possible e.g. in relationship with their energy level. When activated, each node executes Algorithm I . An activated node can be free or cluster-member. In the latter case, it does nothing. In the former case, the activated node tries to create a cluster. To create the cluster, the node tries to calculate a hull around its free and gateway neighbors (gateway nodes can be shared between clusters). In the next paragraph, we explain how this hull is created. If the node succeeds in the hull creation, it becomes the cluster-head of a newly created

191

cluster. Neighbors that are part of the hull or lie inside the hull receive a notification message of the cluster-head about their new status, respectively, gateway node and non-gateway node. The notified nodes broadcast their new status. Nodes outside the hull (the OutsideHulE variable in the algorithm) are not notified. If the creation of the hull fails, nothing happens. Note that after the activation phase, a free node also participates in the sensing and routing tasks of the network, as such, one can look at these nodes as gateway nodes.
Algorithm 1 algorithm of an activated sensor node Require: neighborhood information 1: if m y s t a t e is f7-ee then 2: Hull, OutsideHull crenteHuEl() 3: if Hid1 not InvalidHdl then

Upper hull
A
*

----U

Lower hull

LJ

BestQ

Q ? !

right-to-left scan

Fig. 4. Calculation o f the lower hull

LowerHull. The BcstQ becomes the P node in the next iteration. When no BestQ is found i.e. when all 4: myStute -+clustcrhead Qs are out of transmission range, the algorithm fails. 5: Notify cluster-members about their new state When BestQ is the last node B, the algorithm ends 6: end if successfully. To find nodes that are outside the hull a 7: end if right-to-left scan is easiest (in the example from node B to node P). When node A performs the right-to-left We now discuss the construction of the hull around scan, node S is appended to TrripList because it is the free and the gateway neighbors of an activated node. out of transmission range while node P stays BestQ. The nodes of the hull will become gateway nodes and After all Qs are scanned, the TmpList is added to the will be responsible for routing. To make efficient routing OvtsideHuEl list. The TmpList must be emptied when possible and to guarantee network connectivity, each hull new BestQs are found. node must be able to reach its two neighboring hull nodes Let n be the number of neighbors of a node, then (to create a graph cycle) and the cluster-head must lie the time complexity of the algorithm is as follows: the inside the hull. Due to these requirements and especially node sort can be done in O(nlogn) while the time when the node degree is low, the hull construction might complexity of the while-loop and the inner for-Ioop is fail (the activated node simply stays a free node and can O ( n 2 ) Therefore, the algorithm runs in time O(n2). . later be selected as a (non-)gateway node by another Next, we discuss the results. node). The hull is obtained by combining two parts: the IV. RESULTS lower hull and the upper hull. Algorithm 2 explains in The algorithms were implemented in the Python dispseudo-code how the lower hull is created (upper hull creation is analogous). The algorithm returns the list of crete event simulator SimPy [ 3 8 ] . Figure 5 shows the lower hull nodes (variable LowerHutl) and the list of graphical output of a simulation with 50 nodes and an nodes that are positioned outside the hull because of average node degree of 30. All lines represent wireless connections between communicating nodes. The black connectivity problems (variable Outsidel-lull). The algorithm starts with a sort on x coordinate lines are also cluster borders. Node A is a cluster-head, thereby finding the most extreme nodes (nodes A and node B a gateway node and node C a non-gateway node. B in Figure 4). The while-loop iterates over all nodes Node D is a free node and is not a member of any cluster. starting from node A (lowest x value) to last node B One may notice that the two clusters overlap (arrow). (highest x value). At its turn, node P does a scan over This is due to the sharing of gateway nodes between all Q nodes (see Figure 4) looking for the Q node which clusters. After the upper-right cluster was created firstly, is within transmission range and which maximizes the the lower-let3 cluster includes one gateway node of the P-Q corner. By, maximizing the P-Q corner, a maximal upper-right cluster inside its hull. The overlap can be number of nodes can be included inside the hull. In the removed easily by making this node a non-gateway node. figure, corner U is bigger than corner 6 and therefore, More detailed analysis is needed to decide whether or BestQ is selected to be appended to the Iower hull list not the overlap improves energy efficiency since, on the

-+

192

Aleorithm 2 algorithm to calculate a lower hull Require: Nodes: clusterhead, free neighbors and gateway neighbors I: LowerHulE -% EmptgList 2: OutsideHull + EnzptyList 3: Nodes.sort() 4: P I\ioder.first() 5: Done + Fulse 6: while not Done do 7: if P not cliLsterhead then 8: LowerHuEl.ap~e7bd(P) 9: else IQ; Return InvalidJIull 11: end if 12: BcstQ + Nu11 13: TmpList e EmptyList 14: for all Q's between Nodes.last() and Nodes.indez(P) do 15: if P and Q are in transmission range then 16: if BestQ is Null then 17: BestQ Q 18: else if Comer(P,Q ) > Corner(P,BestQ) then 19: TmpList -eEmptyList
20:

Fig. 5. Example of a simulation with 50 nodes and average node degree of 30

110,

...,..

. , , . .

,,....,.. ,

-.,

BestQ

21: 22:

end if
else if BestQ not Null and Corner(P!Q)> Corner(!', BestQ) then TrnpList .a.ppen,d(Q ) end if end for OutsideHdl .appe-rzd(TmpList) if BestQ is Null then Retum InvaEidHulE else if BestQ is Nodes.East() then if BestQ not clusterhead then LowerHull .append( BestQ) Done T T U ~ else Return InvalidHuEE end if

23: 24:

25:
26:

27:
28:

Fig. 6. Number of cluster-heads in relationship with the average node degree

29: 30: 31: 32:


33:

34: 35: 36: 37: 38: 39: 40:

eIse
P e BestQ end if end while Return LowerHull and OiitsideHuEl

one hand, it provides more and maybe more efficient routing paths, on the other hand, there are less nongateway nodes. In all our simulations we kept the cluster overlap. Note that the gateway nodes and the free nodes together form a connected graph theoretic dominating set. We did run more than 2500 simulations. In each simulation, 1000 sensor nodes are randomly dispersed over the sensor field. The results are shown in Figure 6, 7 , 8 and 9. The error bars display the 95% confidence interval around the mean. Due to the high number of hull failures at very low

193

tm

, , ,

.I

.......................................

.im

** % .

" " " " ' " " ' " " ' " " ' " " ' " " ' ~ " " " " ' ~ " "

nww~dsprse

Fig. 7. Number of hull failures in relationship with the average node degree

Fig. 9. Number of gateway nodes in-relationship with the average node degree

substantial number of non-communicating non-gateways nodes. This number highly depends on the dynamics of the phenomenon and the prediction model. It is clear that the average node degree must be at least high enough to make clustering possible i.e. when hull failures are rare.

v. CONCLUSlON A N D FUTURE WORK


We presented a novel clustering approach for prediction-based monitoring in large-scale, high-density and homogeneous wireless sensor networks. In this
approach, the cluster-heads execute a prediction model

Fig. 8. Number of non-gateway nodes in relationship with the average node degree

node degrees (see Figure 7), the number of clusterheads firstly increases (see Figure 6). After the hull failures become rare i.e. with an average node degree higher than 15, the number of cluster-heads decreases. With increasing node degree, the number of non-gateway nodes increases (see Figure 8) while, complementary, the number of gateway nodes decreases (see Figure 9). In the results, the number of free nodes is added to the number of gateways nodes since they also have a routing task. Energy savings are achieved by turning off the communication radio of the non-gateway nodes whose measurements are predicted correctly by the cluster-head. To compensate for. the overhead of clustering, we need a

while gateway nodes at the circumference of the clusters are responsible for the routing task. Non-gateway nodes that lie inside the clusters are allowed to sleep if they satisfy their cluster-head's predictions. To achieve this cluster infrastructure, an algorithm is proposed and evaluated. Future research includes a detailed analysis to determine at which node degrees clustering is beneficial in terms of energy efficiency. Furthermore, a study and evaluation of prediction models is needed.

REFERENCES
[l] B. W. West, P. G. Flikkema, T. Sisk, and C.W. Koch, "Wireless sensor networks for dense spatio-temporal monitoring of the environment: a case for integrated circuit, system, and network design," IEEE CAS Workshop on Wireless Communications and Networking, University o Notre Dame, August 2001. f [2] S . Til&, N. B. Abu-Ghazaleh, and W. Heinzelman, "A taxonomy of wireless micro-sensor network models," ACM SIGMOBILE Mobile Compuring and Conamunictions Review, vol. 6, no. 2, pp. 28-36, April 2002.

194

[3] C . yee Chong and S. Kumar, Sensor networks: Evolu! tion, opportunities, and challenges, Proceedings of fhe IEEE, vol. 91. no. 8, August 2003. [41 1. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, A survey on sensor networks, IEEE Communications Mogazine, August 2002. [51 R. Szewczyk, E. Osterweil, 1. Polastre, M. Hamilton, A. Mainwaring, and D. Estrin, Habitat monitoring with sensor networks, Communications of the ACM, vol. 47, no. 6, pp. 3440, 2004. [6] H. Chan and A. Perrig, Ace: An emergent algorithm for highly uniform cluster formation, Wireless Sensor Networks, First European Workshop, EWSN 2004. Berlin, Germany? January 19-21, 2004, Proceedings, Lecture Notes in Computer Science 2920 Springer 2004, ISBN 3-540-2U825-9,pp. 154-171, 2004. [7] l Krishna, N. H. Vaidya, M. Chatterjee, and D. Pradhan, A ? cluster-based approach for routing in dynamic networks, ACM SKGCOMM Computer Communicurion Review, vol. 27, no. 2, pp. 49-64, April 1997. [ 8 ] D. G. Thaler and C. V. Ravishankar, Distributed top-down hierarchy construction, Proceedings o the Seventeenth Annual f Joint Conference of the IEEE Computer and Communications Societies (IEEE INFOCOM 1498). San Francisco. California, March 31-April 2, pp. 693-701, 1998. 191 S. Banerjee and S. Khuller, A clustering scheme for hierarchical control in wireless networks, lEEE Injocom 2001, Anchorage, Alaska, April 2001. [IO] H. Zhang and A. Arora, Gs: Scalable self-configuration and f self-healing in wireless networks, Proceedings o the twentyj r s r annual symposium on Principles of distributed computing (PODC 2002). Monterey, California, pp. 58-67, 2002. i l l ] D. Estrin, R. Govindan, J. Heidemann, and S. Kumar, Next century challenges: Scalable coordination in sensor networks, Proceedings o the jifth annunl ACM/IEEE internarianul Conf

ference on Mobile Computing and Nenvorking (MobiCom 1999). pp. 263-270, August 1999. [I21 J. Edgar H. Callaway, Wireless Sensor Netwurks, Architectures and Pmtocols. CRC Press LLC. August 2 W . [13] M. Chattejce, S. K. Das, and D. Turgut, Wca:a weighted clustering algorithm for mobile ad hoc networks, Clusrer Computing, vol. 5, no. 2, pp. 193-204, April 2002. [I41 M. Gerla and J. T. C. Tsai. Multicluster, mobile, multimedia radio network, Wireless networks, vol. 1. no. 3, pp. 255-265, 1995. [lS] C. R. Lin and M. Gerla, A distributed control scheme in multihop packet radio networks for voice/data traffic support, Pmceedings o j IEEE CLOBECOM, pp. 1238-1242, 1995. [16] , A distributed architecture for multimedia in dynamic wireless networks, Proceedings of IEEE GLOBECOM, pp. 1468-1472, 1995. [I71 A. K. Parekh, Selecting routers in ad-hoc wireless networks, Proceedings of ihe SBT/IEEE Internatiunal Telecommunications Symposium, August 1994. [IS] D. J. Baker and A. Ephremides, A distributed algorithm for organizing mobile radio telecommunication networks, Proceedings of the second international conference on diswibuted computer systems, pp. 476483, April 1981. 1191 The architectural organization of a mobile radio network via a distributed algorithm, IEEE Transactions on Cnmmunicarinnr COM, vol. 29, no. 11, pp. 1694-1701, 1981. [lo] A. Ephremides, . . Wieselthier, and D.J. Baker, A design I E. concept for reliable mobile radio networks with frequency

-.

hopping signaling, Proceedings of IEEE, vol. 75, no. 1, pp. 5673, 1987. {Zl] S. Basagni, Distributed clustering in ad hoc networks. Proceedings of Internarional Symposium on Parallel Archirectures, Algorithms and Networks, pp. 31@315, June 1999. [221 , Distributed and mobility-adaptive clustering for multimedia support in multi-hop wireIess networks, Proceedings o Vehicular Technobgy Conference, vol. 2, pp. 889-893, Fall f 1999. [23] Q. Fang, E Zhao, and L. Guibas, Lightweight sensing and communication protocols for target enumeration and aggregation, Proceedings of the fourth ACM International Symposium an Mobile Ad-hoc Nerworking and Computing (MobiHocj, pp. 165-176, 2003. [24] W. R. Heinzelman, A. Chaodmkasan, and H. Balaknshnan, Energy-efficient communication protocol for wireless microsensor networks, Proceedings of the 33rd Hawaii International Conference on System Sciences, Maui, Hawaii, vol. 8, January 2000. [25] 3. N. Clark, C. J. Colbourn, and D. S. Johnson, Unit disk graphs, Discrete Mathematics, vol. 86, p p , 165-177, 1990. [26] L. Jia, R. Rajaraman, and T. Suel, An efficient distributed algorithm for constructing small dominating sets, Distributed Computing, Special issue: Selected papers fmm PODC 01, vol. 15, no. 4, pp. 193-205, 2002. [27] E Kuhn and R. Wattenhofer, Constan-time distributed dominating set approximation, Proceedings of fhe twenty-second annual symposium on Principles of distributed computing, Boston, Massachusetts, pp. 25-32, 2003. [28] B. Das, R. Sivakumar, and V. Bhargbavan, Routing in ad hoc networks using a spine, Proceedings of the 6th Intemntionnl Conference on Computer Communications and Networks, p. 34, 1997. [2Y] K. Alzoubi, P.-J. Wan, and 0. Frieder, New distributed algorithm for connected dominating set in wireless ad hoc networks, big island, hawaii, 35th Annuai Hawaii Inremational Conference on System Sciences (HICSSOZ), 9, January 2002. vol. E301 J. Wu and H. Li, On calculating connected dominating set for efficient routing in ad hoc wireless networks, Proceedings o the 3rd intemtional workshop on Discreie algorithms and f methods for mobile computing and communications, Seattle, Washington, United States, pp. 7-14, 1999. [31] J. Wu, Extended dominating-set-based routing in ad hoc wireless networks with unidirectional links,-IEEE Trunsocrions on Parallel and Disrribured Computing, vol. 22, no. 1 4 , pp. 327340, 2002. [32] I. Carte and D. Simplot-RyL Energy-efficient area monitoring for sensor networks, Computer, vol. 37, no. 2, pp. 40-46, 2004. 1331 E Dai and J. Wu, Distributed dominant pruning in ad hoc networks, Proceedings of the IEEE International conference on Communications (ECC 2003). IEEE Press, pp. 353-357, 2003. [34] S. Goel and T. Imielinski, Prediction-based monitoring in sensor networks: Taking lessons from mpeg, ACM Computer Communication Review, vol. 31, no. 5 , October 2001. [35] Y Xu and W.-C. Lee, On localized prediction for power efficient object tracking in sensor networks, 23rd Inrernational Conference on Distributed Computing Systems Workshops (ICDCSW03). Providence, Rhode Island, USA, May 2003. [36] Y.Xu. J. Winter, and W.-C. Lee, Prediction-based strategies for energy saving in object tracking sensor networks, IEEE Inter-

195

national Conference on Mobile Datu Management (MDMLM), Berkeley, Calif~mia. January 2004. [37] N. Bulusu, J. Heidemann, and D. Estrin, Gps-less low cost outdoor localization for very small devices, University of Southern California, Tech. Rep. 00-729, April 2000.

[SS] Si*.

[Online]. Available: http://simpy.sourcefor~e.nnetl

196

Вам также может понравиться