Distributed direction-based localization in wireless sensor ... - CiteSeerX

Computer Communications 30 (2007) 1424–1439 www.elsevier.com/locate/comcom

Distributed direction-based localization in wireless sensor networks Sheng-Shih Wang b

a,*

, Kuei-Ping Shih b, Chih-Yung Chang

b

a Department of Information Network Technology, Chihlee Institute of Technology, Banciao 220, Taipei, Taiwan Department of Computer Science and Information Engineering, Tamkang University, Tamshui 251, Taipei, Taiwan

Received 7 August 2006; received in revised form 3 January 2007; accepted 10 January 2007 Available online 17 January 2007

Abstract Location awareness is an attractive research issue in the wireless sensor network (WSN). However, precise location information may be unavailable due to the constraint in energy, computation, or terrain. Additionally, several applications can tolerate the diverse level of inaccuracy in such geographic information. Thus, this paper presents a direction-based localization scheme, DLS, whose main goal is for each sensor to determine its direction rather than its absolute position. The direction we are concerned with is the one relative to the sink. Motivated by the proposed spatial locality property, DLS considers multiple messages received for a sensor to determine its direction. Furthermore, a novel scheme, anchor deployment strategy, is also proposed for the improvement of the estimated correctness in direction of the sensor within the communication range of the sink. With the aid of the virtual dual direction coordinate (VDDC) system, DLS is able to efficiently and precisely position sensors around the axes. We evaluate DLS via simulations in terms of various numbers of sensors and communication ranges for the scenarios with different numbers of directions. The average correct rates in DLS reach approximately 94%, 86%, and 81% for the networks with 4, 8, and 16 directions, respectively. DLS achieves outstanding performance for the high density networks as well. In addition, DLS also works well regardless of the sink placement. Overall, simulation results validate the practicality of DLS, and show that DLS can effectively achieve direction estimation.  2007 Elsevier B.V. All rights reserved. Keywords: Wireless sensor networks (WSNs); Localization; Range-free; Direction-based localization scheme (DLS)

1. Introduction Recently, the remarkable advance in Micro-ElectroMechanical Systems (MEMS) has enabled the evolution of the tiny wireless device, called sensor, which is characterized by its low cost, low power, short radio range, and multiple functions [1]. A large number of sensors deployed in either the ad hoc or pre-scheduled manner form a wireless sensor network (WSN). The WSN is now in widespread use for a variety of applications, including object detection, target tracking, security surveillance, and environmental monitoring [2–5]. Essentially, sensors are always arbitrarily * Corresponding author. Tel.: +886 2 2257 6167x601; fax: +886 2 2255 4448. E-mail addresses: [email protected] (S.-S. Wang), kpshih@mail. tku.edu.tw (K.-P. Shih), [email protected] (C.-Y. Chang).

0140-3664/$ - see front matter  2007 Elsevier B.V. All rights reserved. doi:10.1016/j.comcom.2007.01.002

scattered in the sensor field without geographic information known in advance. However, numerous applications rely on the geographic information (e.g., location) of the sensor to identify the position of the tracking object [4,6], to assist in delivering packets to the fields of interests [7,8], to reduce the number of packets flooding the network for route discovery [9,10], and to provide sensor deployment for mitigating coverage overlap [11,12]. Localization, for a sensor to determine its location information has become an attractive research issue in WSNs. Much previous research has proposed numerous localization schemes, which are generally classified into rangebased and range-free localization schemes, depending on the use of distance (range) or angle estimate [13–16]. GPS is a wide-area system for sensor localization, but extremely expensive and energy-consuming properties make it impractical to be installed in a sensor [17]. Other range-

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

based schemes, including RSSI, AoA, ToA/TDoA, and several protocols based on these representative mechanisms, are proposed in the literature [18–21]. These approaches achieve sensor positioning, but have constraints in hardware cost (e.g., GPS receiver or smart/directional antenna) and time synchronization. Unlike the range-based technique, the range-free scheme enables sensors to learn their location information without the aid of range estimates [13,14,22]. Such techniques, like APIT [14], and MDS [15], DV-hop [23] generally require numerous location-aware nodes, by which locationunknown sensors are able to determine their locations. Comprehensive simulation results implemented in the literature have confirmed that compared with the range-based solution, the range-free mechanism is suitable for sensor positioning due to its cost-effectiveness [14,24]. The majority of existing localization schemes focus on sensor coordinate estimation. However, the precise position of the sensor is probably difficult to obtain owing to the constraint in energy, computation, or terrain. In WSNs, several applications can tolerate diverse levels of inaccuracy in location, depending on their requirements [25]. Consider a coarse-grained coordinate system represented by the direction and the hop distance related to the sink, if the location of a forest fire is identified by the aforementioned geographic information, then the rescuers can be aware of the approximate scene of such fire according to the direction and the hop distance related to them. When the rescuer takes action to go to the disaster region for fire extinguish, a sensor will provide the geographic information to the rescuer within the communication range of the sensor. Obviously, by means of the information, the rescuers realize whether they are in the accurate direction, and are aware of the distance between the position they locate and the destination as well. The rescuers are most likely to be smoothly guided towards the fire scene. As a result, we reasonably conclude that direction and distance information of sensors are sufficient and favorable for a variety of applications, especially for ones without the requirement of high level of accuracy in location information. The paper develops a fully distributed localization scheme, DLS, which enables a sensor to estimate its direction without GPS-support. Sensor direction mentioned in the paper means the relative one to the sink. The direction is represented in the form of the gray code, by which a sensor effortlessly enables to recognize whether a packet comes from one of its adjacent directions or not. The main idea of DLS is the spatial locality property, by which all sensors are able to determine their directions according to the packets received. Additionally, three novel mechanisms, anchor deployment strategy, multi-message decision scheme, and virtual dual direction coordinate system, are introduced in DLS to assist the sensor in improving the estimated correctness. To our best knowledge, this study is the first investigation to concentrate on direction estimation of the sensor. DLS involves the following significant advanta-

1425

ges: (1) DLS is a fully distributed localization scheme. (2) DLS achieves well accurate direction estimates. (3) DLS is cost-effective for no need of additional facilities (e.g., GPS receiver). (4) DLS does not require complicated computation for sensor direction estimation. The rest of this paper is organized as follows. Section 2 describes the related works in terms of localization in WSNs. Section 3 formulates our network model, and gives an overview of DLS. Section 4 then mentions a significant property, spatial locality property. Next, a novel localization scheme based on sensor direction, DLS, is proposed in Section 5. The extensive analyses and discussions of the effect of sink placement and physical location ambiguity are presented in Section 6. Meanwhile, the simulation results are shown in Section 7. Finally, Section 8 concludes the paper and presents future research directions. 2. Related works Various localization approaches have been proposed in the literature. These schemes are categorized as rangebased and range-free, depending on the use of distance or angle estimate. 2.1. Range-based localization schemes GPS, a representative positioning device, is widely used in localization due to its simplicity [2,26]. Based on radio waves and time spacing, GPS can exactly locate the nodes with the limited error, especially for the outdoor environments. Two time-based methods, time of arrival (ToA) and time difference of arrival (TDoA), are used in [17,18]. ToA is based on the range estimations by the signal arrival time, while TDoA relies on the difference in time between two arrived signals. Angle of arrival (AoA) to estimate node position is proposed in [19]. The principle of AoA is to detect the originators of signals according to the angle of arrival, and then to calculate the positions of nodes by means of triangulation. Although achieving high estimated correctness, these approaches require each sensor node to be equipped with the additional facility, such as smart/ directional antennas, antenna array, or ultrasound receiver. Obviously, expensive hardware cost makes these solutions impractical to WSNs. Maximum likelihood estimation (MLE) is an alternative used in AHLoS system (Ad-Hoc Localization System) [17], whose aim is to minimize the differences between the measured distances and estimated distances to determine the positions of nodes. An infrastructure-free positioning algorithm without the aid of GPS is proposed in [20]. The protocol focuses on constructing a relative coordination system to compute node positions, but requires numerous message exchanges in the course of system construction. Besides, various RF-based and ultrasonic location systems such as RADAR [21] and Cricket [18] are based on the principle of trilateration to measure the

1426

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

distance between two nodes. A distributed and scalable GPS-free localization mechanism, based on the clustering technique, is mentioned in [27]. The approach first requires the formation of local coordinate systems at a subset of nodes, and then establishes a global coordination system for the entire network based on the local coordinate systems. Such cluster-based method generates less communication overhead, but it has the drawback in heavy computation. A trilateration-related localization scheme is proposed in [28]. The scheme based on the AoA estimation technique enables a sensor to determine its location by listening to messages from three or more stationary beacon nodes (BNs). To guarantee that all sensors are able to receive the messages from BNs, each BN has to be equipped with special transmission facility, such as smart antenna or directional antenna. Such transmission capability obviously requires additional hardware cost. Therefore, the scheme is not cost-effective. Additionally, for some applications, the environmental conditions such as fading, obstacles, and multipath reflections are most likely to degrade the accuracy in location since all sensors have to receive the beacon signals from at least three BNs. 2.2. Range-free localization schemes Aside from the mechanisms relying on distance measurements, numerous alternative solutions, mainly categorized as the range-free schemes, have also been considered in [14,15,23,24,29]. The scheme requires a small percentage of location-aware nodes (e.g., anchors or beacons) for sensor positioning. In principle, a sensor mainly calculates its position depending on the anchor position in the packet received. Additionally, the hop count (namely, the number of hops between a sensor and an anchor) is also employed to improve the accuracy in sensor localization. An amorphous positioning algorithm, DV-Hop, addressed in [23], employs offline hop-distance estimations to improve location estimates via neighboring information exchanges. In [14], an area-based scheme, called APIT, accurately performs location estimations especially in sensor networks with the irregular radio pattern and random node placement. The main idea of APIT is to divide the whole network into triangular regions among anchors, and then to determine the possible position where a sensor locates via the aggregation of the two distinct triangular regions. Consequently, the position of a sensor can be estimated by calculating the center of gravity (COG) of the intersections of the triangles where a sensor resides. Through the simulations, APIT not only reduces the system cost, but also possesses the reasonable accuracy in comparison with many costly solutions, including Centroid [13], DV-Hop [23], and Amorphous [29] schemes. A method inspired by a data analysis technique, multidimensional scaling (MDS), is proposed in [15]. The algorithm, called MDS-MAP starts with rough estimations of distance between each pair of sensors by means of the

all-pairs shortest-paths algorithm. Classical MDS is then used to derive node locations. Based on classical MDS, a novel iterative MDS technique presented in [24] is exploited to compute the relative positions of adjacent sensors. Through transformations such as scaling, rotation, and reflection, the relative positions of sensors are consequently aligned to map physical coordinates with the aid of three or more location-aware anchors. 2.3. Summary Basically, DLS focuses on the sensor direction related to the sink instead of the absolute physical position. All sensors are able to determine their directions depending on the packets received. DLS is a range-free localization scheme for no need of the dependence on the distance information. Compared to the existing approaches, DLS does not require the sensor equipping with the additional facility, including GPS receiver or smart antennas. Unlike some localization schemes [15,27], which need to first construct the relative map, and then perform the geometric transformations (e.g., scaling, rotation, and reflection), DLS enables each sensor to determine its direction without requiring any geometric transformation. Significantly, DLS is fully distributed because it depends on the local information to achieve sensor localization. Additionally, a sensor is able to quickly determine its direction relying on the packets received. 3. Preliminaries Localization in WSNs can be regarded as the location discovery problem. Unlike much research, whose main objective is precise coordinate estimation, the localization problem we are concerned with is that given a sink and some direction-aware anchors, each sensor is able to determine the direction relative to the sink and the hop distance from the sink. 3.1. Network model The network considered is a square area. Without loss of generality, the sink is placed at the center. We are given Ns stationary sensors, si, i = 1, 2, . . . , Ns, with unknown directions and positions. The sensor is termed unknown sensor. All unknown sensors are uniformly scattered in the network. The communication range of a sensor, denoted as r, is a circle centered at the sensor. Each sensor has the communication capability, so as to exchange messages. Currently, we consider an obstacle-free environment, in which each sensor is able to communicate with all of its neighbors. We also consider a connected network, within which each sensor has at least one neighbor. Let Ndir = 2n be the number of directions, where n is a positive integer, which is determined in advance. Given Ndir = 2n directions, the network is virtually partitioned into 2 · Ndir distinct regions, comprising Ndir directional

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

regions and Ndir axis regions. Each axis region is divided into two sub-regions of equal size with a common virtual borderline called an axis. Fig. 1 shows an example of our network model. Apparently, a sensor within an axis region is able to receive the packets from sensors at both the same and the neighboring directions. Two virtual direction coordinate systems, virtual primary direction coordinate (VPDC) system and virtual auxiliary direction coordinate (VADC) system, are created to represent the direction of each sensor and to correct the position of the sensor near the axis, respectively. The VPDC system is expressed by the primary direction code, while the VADC system is represented by the auxiliary direction code. The primary and auxiliary direction codes of sensor si are, respectively, denoted by cpri (si) and caux(si). Sensor direction in the paper means the one in the VPDC system (i.e., primary direction code), so DLS focuses on the estimation of such code of each sensor. Both direction codes are represented in the form of the gray code, arranged counterclockwise and numbered in order. The gray code is a method of encoding binary numbers which has the property that two consecutive numbers differ only in one bit. Based on the gray code representation, each sensor can effortlessly realize whether a packet comes from one of its adjacent directions or not. Most previous works mentioned the benefit of locationknown sensors placement in localization [13,14]. DLS, thus, deploys a small fraction of direction-aware sensors, called anchors, via either digital compasses or manual presetting. That is, each anchor is aware of its direction. The numbers of anchors required in DLS is the same as the number of directions (namely, Ndir). All anchors are evenly placed at the range r from the sink on each axis (i.e., there exists only one anchor on each axis).

(2)

r (1)

A3

(1)

A2 A2

R 2 (01)

R 1 (00)

1427

The communication ranges of the sink, anchors, and unknown sensors are assumed to be identical. DLS exploits the radio propagation model in [30], by which the receiver is able to measure the signal strength in spite of attenuation of radio signal. The perfect spherical radio propagation is also assumed. Definition 1. Given two sensors si and sj, sj is said to be the neighbor of si if sj is within si’s communication range, and vice versa. Recall that we use the gray code representation to stand for the direction code. The code assignment of the anchor follows the counterclockwise manner, by which anchor ai + 1 is regarded as in the counterclockwise direction of anchor ai. In Fig. 1, four anchors (namely, a1, a2, a3, and a4), whose primary direction codes cpri(a1), cpri(a2), cpri(a3), and cpri(a4) are, respectively, indexed by 00, 01, 11, and 10, are placed on each axis. The anchors associated with axes virtually partition the network into 8 regions, termed Ri and Ai, for i = 1, 2, . . . , 4. The sub-regions of axis region ð1Þ ð2Þ Ai are denoted as Ai and Ai . Actually, the primary direction code of each directional region is assigned by the sink. Suppose the direction code of Ri and Ai are, respectively, denoted by cpri(Ri) and cpri(Ai). As Fig. 1 shows, the primary direction codes of four directional regions, cpri(R1), cpri(R2), cpri(R3), and cpri(R4), are regarded as 00, 01, 11, and 10, respectively. The direction codes of axis regions A1, A2, A3, and A4 are regarded as 00, 01, 11, and 10, respectively. All sensors in a certain directional or axis region have the same primary direction code. For example, ð1Þ ð2Þ all sensors within R1, A1 , and A1 share direction code 00. 3.2. Basic concepts Fig. 2 depicts an overview of DLS. Suppose the network has a high density of sensors. The communication ranges of both the sink and sensors are assumed to be unit disks. All sensors are arbitrarily distributed in the network. For ease of explanation, the sink is placed at the center of the net-

r

01

a2 a3

a1

(2)

A3

00

(2)

r

A1

r

A1

(1)

a4 R 3 (11)

axis

(1)

(2)

A4 A4

sink

R 4 (10)

anchor

Fig. 1. Example of the network model with Ndir= 4. Without loss of generality, the sink is placed at the center. Four anchors (i.e., a1, a2, a3, and a4) are deployed at the range r to the sink on each axis. The network is virtually partitioned into 8 regions: 4 directional regions (i.e., R1, R2, R3, and R4) and 4 axis regions (i.e., A1, A2, A3, and A4). Each axis region is divided into two sub-regions. Namely, axis region Ai includes sub-regions ð1Þ ð2Þ Ai and Ai .

11

10 sink

anchor

sensor

Fig. 2. Overview of DLS. The sink originates an LREQ packet outward, and the packet then floods throughout the network by means of packet dissemination. All unknown sensors will determine their directions in accordance with the LREQ packets received.

1428

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

work. Ndir is assumed to be 4 so that each region is represented by the direction code with only 2 bits. All anchors evenly lie on each axis at the distance r from the sink. DLS performs after sensors are deployed. Inspired by the spatial locality property addressed in Section 4, DLS achieves sensor direction estimating according to the idea of packet dissemination, by which an unknown sensor is able to estimate its direction based on the direction of its neighbors. A sensor can realize such information carried in the received packet. The request of sensor localization is originated by the sink, so DLS intends to identify the directions of the sensors near the sink as correct as possible. To achieve the goal, DLS applies the anchor deployment strategy, whose main idea is locating all anchors near the sink. Because the anchor is identified as the exact direction, the anchor deployment strategy significantly guarantees that sensors within the communication range of the sink are able to obtain the correct directions, especially for large Ndir. The property will be discussed in Section 5. DLS is initiated by the sink, which issues a Locating Request (LREQ) packet to all of its neighbors. The unknown sensors within the communication range of the sink will discard the LREQ packet due to no direction information in the packet. Among the neighbors of the sink, only anchors require sending the LREQ packets with their own directions. Once receiving an LREQ packet, an unknown sensor estimates its direction according to the direction information involved in the LREQ packet, and then propagates the direction information to its neighbors. The process terminates for a sensor when its direction never changes. Recall that a sensor determines its direction according to the direction in the packet received. Upon the receipt of a packet, a sensor invokes direction estimating. The procedure is called the one-message decision scheme. Although achieving direction estimating quickly, however, such scheme probably leads to erroneous estimate, especially

4. Spatial locality property As mentioned before, packet dissemination enables a sensor to determine its direction by means of the information in the received LREQ packet(s). Namely, the estimate for a sensor is obviously associated with the direction of its neighbors. Thus, we made numerous prior investigations to observe the impacts of the packets received at a sensor. Fig. 3 shows the result in terms of different numbers of sensors and communication ranges. For ease of explanation, the directions where LREQ packets come from are categorized as the same, adjacent, and other, which are listed as below. • Same: Given a sensor si with direction code cpri(si), if a sensor sj is si’s neighbor and cpri(si) = cpri(sj), then for sensor si the packet sent from sj is indicated by the same. • Adjacent: Given a sensor si with direction code cpri(si), if a sensor sj is si’s neighbor, and the difference between cpri(si) and cpri(si) is only in one bit, then for sensor si the packet from sj is indicated by the adjacent. • Other: For a sensor si, the packet from neither the same nor the adjacent is denoted by the other. Consider a network with Ndir = 4, for a sensor with direction code 00, a packet from direction 00 is indicated by the same, while a packet from either direction 01 or 10 is viewed as the adjacent. A packet originated from

b 100 90 80 70 60 50 40 30 20 same

10

adjacent

other

0 500

600

700

800

Number of sensors

900

1000

Percentage of received packets (%)

Percentage of received packets (%)

a

for a sensor located near the axis. Therefore, in DLS, we devise a multi-message decision scheme, by which a sensor takes a reasonable duration to gather more packets sent from its neighbors once receiving an LREQ packet. The approach really improves the accuracy in sensor direction estimating. This section formulates the network model and addresses the basic concepts of DLS. Meanwhile, the subsequent sections will introduce the spatial locality property and mention DLS in detail.

100 90 80 70 60 50 40 30 20 same

10

adjacent

other

0 50

60

70

80

Communication range (m)

Fig. 3. Spatial locality property. The network is a 500m · 500m square region with Ndir = 4. The sink is at the center of the network. (a) Percentages of received packets from different directions for a sensor vs. number of sensors. (b) Percentages of received packets from different directions for a sensor vs. communication range.

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

the direction other than 00, 01, and 10 is represented by the other. The observation in Fig. 3 reveals that for a sensor, approximately 88.8% packets received come from the sensor(s) in the same direction. The percentages of packets from either the adjacent or other region reach about 5.1% or 6.1%, respectively. In general, a sensor close to the axis most probably receives packets from the sensors in the same or the adjacent direction. Besides, a sensor within the communication range of the sink may receive more packets from the sensors in the other directions than the one near the axis. On the other hand, a sensor will not receive any packet from the adjacent and the other directions if it is far from the axis. Significantly, the increase of the number of sensors will result in more sensors which only receive the packets from the same direction. However, the number of sensors either close to the axis or within the communication range of the sink increases as well. Thus, in Fig. 3(a), the average percentage of received packets from either the same, the adjacent, or the other direction approximately remains identical even if the number of sensors increases. Based on the above result, we conclude that a sensor most probably receives the packet(s) from other one(s) located in the same direction with regardless the number of sensors in the network as well as the communication range, and formulate the following spatial locality property. Property 1. Spatial locality is that most of the packets received at a sensor are likely to be delivered from its neighbors located in the same direction. According to the spatial locality property, a sensor and its neighbors are most likely to have the same direction code. Recall that DLS uses the direction information to determine a sensor’s direction. An unknown sensor si, thus, requires maintaining the information in its neighboring table, each of which is a four-tuple with the format (id(sj), cpri(sj), caux(sj), hop(sj)), where sj is one of si’s neighbors, id(sj) stands for the identifier of sj, cpri(sj) and caux(sj) denote sj’s primary and auxiliary direction codes, respectively, and hop(sj) is the minimum hop distance from sj to the sink.

1429

DLS requires a small percentage of direction-aware anchors as well. With the inherence of the manual placement, the anchor is able to be reasonably set at the designated position. Since packet dissemination starts from the sink, the idea that enhancing the accuracy in direction of the sensor closed to the sink inspires the investigation to design a refined scheme for direction estimating. Here, we define the critical region as below. The sensor within the critical region is called the critical sensor. Definition 2. Given a sink, the critical region, CR, is the area within which all sensors are able to directly communicate with the sink. Fig. 4 shows a network with Ndir = 8. The direction code used in the VPDC system obviously requires 3 bits. Consider all anchors are evenly placed at the communication ranges from the sink on each axis. For ease of explanation, we, here, focus on region R1, with cpri(R1) = 000. Suppose CR in R1 is divided into CR1 and CR2. A sensor within either CR1 or CR2 is obviously able to receive the LREQ packet from at least two anchors. Namely, all sensors in CR2 are within the communication ranges of anchors a1, a2, and a3, while a sensor within CR1 is able to receive the packet issued from anchors a1 or a2 but not from anchor a3. Note that, some sensors within either CR1 or CR2 are within the communication range of anchor a8. Significantly, a1 and a2 are the two closest anchors to the sensor. We consequently give Rule 1 for code assignment of the critical sensor. Rule 1. Suppose ai and aj are the two closest anchors to the critical sensor si. The direction code of critical sensor si is defined as cpri(ai) if cpri(ai) is precedent to cpri(aj) in the gray code sequence. Obviously, a sensor within either CR1 or CR2 has the accurate direction code 000. All critical sensors in other directions are also aware of the correct direction codes in accordance with Rule 1. Recall that accurate estimation of critical sensors significantly assists unknown sensors in

5. Direction-based localization scheme (DLS) In this section, we propose a fully distributed directionbased localization scheme (DLS), for an unknown sensor to determine its direction. DLS comprises two major components, anchor deployment strategy and multi-message decision scheme, to increase the estimated correctness in direction of the unknown sensor. In addition, a hybrid direction coordinate system, involving the VPDC and VADC systems is introduced to resolve the location ambiguity problem.

a3

a2

a4

CR

R1 (000)

CR2 CR1

a5

a1

a8

a6 a7

5.1. Anchor deployment strategy sink

Numerous existing localization protocols exploit some location-aware anchors to benefit sensor positioning.

anchor

Fig. 4. Anchor deployment strategy. The shaded region in the network is the critical region.

1430

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

direction estimations because the LREQ packet is initiated from the sink. As a result, DLS exploits an elegant anchor deployment strategy, which locates all anchors evenly at the communication ranges from the sink on each axis, to reduce the propagation errors in the course of packet dissemination. 5.2. Multi-message decision scheme As mentioned before, one-message decision scheme only relies on one LREQ packet for sensor direction learning. The localization protocol based on such scheme is straightforward and cost-effective, as well as easy to implement. However, the technique obviously incurs erroneous estimates if a sensor only receives one LREQ packet which comes from the other direction. Inspired by the spatial locality property, a novel multi-message decision scheme is utilized in DLS. The main idea of the scheme is taking a reasonable duration into account to consider more LREQ packets for the improvement of the accuracy in direction estimating. Fig. 5 illustrates the transitions among the states. Each sensor maintains the information of the neighbors in its own neighboring table. All sensors are initially in the Idle states. A sensor in the Waiting or the Learning states is to gather more LREQ packets or to determine its direction according to the received LREQ packets, respectively. Algorithm 1 shows how DLS works in the Waiting and Learning states. In the Sending state, a sensor immediately sends an LREQ packet to propagate the current estimate to all of its neighbors, and then enters the Idle state. 5.3. Virtual dual direction coordinate (VDDC) system The spatial locality property implies that for a sensor most of the received LREQ packets come from its neighbors in the same direction. Although the anchor deployment strategy guarantees accurate estimates for all critical sensors, erroneous estimates are likely to occur during packet dissemination. The circumstance is significantly revealed in the axis region. As Fig. 6 shows, the shaded regions are the axis regions. Each direction has two axis sub-regions. In spite of the utilization of the multi-message decision scheme, a sensor within the axis region may significantly obtain the errone-

send packet

Idle

receive packet

di

re

cti

on

ch

an

ge

d

Waiting

timer timeout

Algorithm 1 DLS – the direction-based localization scheme {In the Waiting state.} Invoke a waiting timer Tw; while the timer Tw is not expired do if receive an LREQ packet then if the sender is the sink then Set critical sensor flag; else if the sender id is present in my neighboring table then Update the entry associated with the received LREQ packet in my neighboring table; else Create a new entry for the received LREQ packet; end if end if end if end while {In the Learning state.} if critical sensor flag is set then Set my direction according to Rule 1; /* anchor deployment strategy */ else Calculate the numbers of entries in my neighboring table for all directions; Set my direction as the direction with the maximum value among all directions; end if if my direction changes then Enter the Sending state; else Enter the Idle state; end if

Sending

direction changed

un

ous estimate. Specially, the direction code of such sensor is probably identical to the one of the adjacent directions. ð2Þ For example, a sensor, si, actually locates within A1 (i.e., cpri(si) = 00), but may be regarded as locating in R4 (i.e., cpri(si) = 10) if most of LREQ packets received come from direction 10. Such incorrect estimation obviously incurs less performance in sensor localization. To efficiently improve the DLS of estimated correctness in direction of the sensor within the axis region, DLS considers an auxiliary coordinate system, VADC, associated with the VPDC system. The VADC system is generated by counterclockwise rotating the VPDC system with an angle a ¼ Npdir . The approach to code assignment in the VADC system resembles that in the VPDC one. As shown in Fig. 7, a system, involving the VPDC and VADC systems, is named the virtual dual direction coordinate (VDDC) system. The dark and gray broken lines, respectively, indicate the VPDC and VADC systems. Figs. 7(a) and (b) show the VADC systems for Ndir = 4 and 8, respectively.

Learning

Fig. 5. State transition diagram of DLS at each sensor.

Fig. 6 demonstrates an example of the coding system exploited in DLS. For ease of description, each direction is assumed to comprise 4 sub-regions. Two of the subregions are directional sub-regions, and the other two ones are axis sub-regions. All sensors hold two direction codes at the end of direction estimation. The main goal of the VADC system is to enable a sensor within the axis region to be aware of the axis to which it is

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

direction 00 (1)

R2 (2)

R2

(2)

(2) A2

r

(1) 1 (2) 1

R1

(1) A2

(1)

R1

r

(1)

r

A1

(2)

r

A1

A3 A3

(2)

(1)

A : (00,10)

A 2 : (01,00)

(2) 1

(2)

(1) A4

(2) A4

direction 10

A : (11,01) (1 )

R 3 : (11,01)

R4

(2)

direction 11 (2) 3

(1)

R3

(1)

R 2 : (01,00) R 2 : (01,01)

A 3 : (11,01)

(2) R4

(2)

R : (00,00)

(1)

(1) R3

(1)

A 2 : (01,00)

(1)

si

direction 01

A : (00,10)

R 1 : (00,10) codi ng

1431

(2)

R 3 : (11,11)

(1)

A 4 : (10,11) (2)

A 4 : (10,11) (1)

R 4 : (10,11) (2)

R 4 : (10,10)

Fig. 6. Coding system with Ndir = 4 in DLS, where r is the communication range of a sensor. Each directional region Ri comprises many regions, ð1Þ ð2Þ ð2Þ ð1Þ individually represented by Ri , Ri , Ai , Aiþ1 and the critical region in Ri. All unknown sensors are identified as two direction codes, represented in the form (primary direction code, auxiliary direction code).

a

b

00 01

01 11

011 001

00 10

10

VPDC system

010

000

110

100

11

111 101

VADC system

VPDC system

010

000 100

110

111

101

VADC system

α=π/4

VDDC system

001

011

α=π/8

VDDC system

Fig. 7. Example of the virtual dual direction coordinate (VDDC) system. (a) VDDC system for Ndir = 4. (b) VDDC system for Ndir = 8.

close although obtaining an incorrect primary direction code. Taking the VADC system into account, each sensor in the network will have two direction codes, primary direction code and auxiliary direction code. The effectiveness of the VADC system is mentioned in the following. In Fig. 7, suppose most of LREQ packets si received come from R4, cpri(si) is determined as 10 according to the multimessage decision scheme. Apparently, the direction code of si is incorrect in case of the only usage of the VPDC system. Based on the spatial locality property, si is most likely to obtain auxiliary direction code 10 because most LREQ ð1Þ ð2Þ ð1Þ ð2Þ packets come from regions R1 , R1 , R1 , and R4 , all of whose auxiliary direction codes are 10. Note that sensor si is most unlikely to receive the LREQ packets from direction 11, so the sensor realizes that it is near the axis 00 rather than axis 10 though cpri(si) = 10. Significantly, the VDDC system efficiently identifies the exact axis to which

a sensor close in spite of the incorrect estimate in the primary direction code. 6. Analyses and discussions In the section, we discuss the robustness of DLS for applications with different sink placements. Then, we further state the physical location ambiguity problem, in which two sensors are likely to locate in the distinct physical positions although holding the same direction code and hop distance from the sink. 6.1. Location of the sink Without loss of generality, DLS assumes the sink is located at the center of the network. One network architecture, whose sink is placed at the edge, is also in wide-spread

1432

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

use for several applications. Such networks revealed in Figs. 8(a) and (b) can be classified into two scenarios, in which the sink is placed either at the corner (Fig. 8(a)), or the edge (Fig. 8(b)) rather than at the center or inside the network. In Fig. 8(a), the sink is aware of its own position (i.e., at the corner of the network) by means of the manual setting, so it has the knowledge that the network is partitioned into two directions. That is, only two direction codes are considered for all sensors in the network. Obviously, only three anchors, whose direction codes assigned by the sink are 000, 101, and 100 are located at the axes. Similarly, totally four direction codes are considered for the scenario shown in Fig. 8(b). In essence, DLS still works as usual for sensor nodes within all directions without requiring any modification and keeps the performance of sensor localization. Another instance of sink placement is demonstrated in Fig. 8(c), where the sink is within the network, but not at the center. Compared to the network mentioned in Section 5, such network virtually comprises many directions with different sizes. However, the network obviously has no variation in number of directions. Because the region size and the number of directions both have no effect on DLS, DLS can be also well performed in the network. Consequently, DLS is practical and well-suited for the network regardless of the type of sink placement. 6.2. Physical location ambiguity DLS advantages each sensor to determine its relative direction and hop distance from the sink. However, DLS may cause the problem that two sensors locate in the distinct physical positions although holding the same direction code and hops. Such two sensors are called ambiguous sensors. Here, let the network density termed D be defined as the number of neighbors per sensor. Assumption 1. Consider two sensors at minimum hop distance h, there exist two values lh and uh such that the Euclidean distance d between such two sensors is bounded. Namely, lh 6 d 6 uh. The quality of the bounds depends on the network density. In particular, for each h > 0, lim uh  lh ¼ r

D!1

ð1Þ

where r is the communication range of a sensor. Basically, Assumption 1 guarantees that a sensor u has at least one neighbor with distance r from sensor u if D ! 1. If Assumption 1 holds, we give a lemma to show the maximum hop distance between two ambiguous sensors. Lemma 1. Consider a network with Ndir = 2n for n > 0, and the sink S is at the center. The maximum hop distance c between two ambiguous sensors at hop distance h from the sink is determined as follows.

 c¼

2h if n ¼ 1 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2ð1  cosaÞ if n > 1

h

ð2Þ

is the angle formed by two neighboring axes where a ¼ 2p 2n intersecting at S. Proof. In Fig. 9, suppose two ambiguous sensors A and A 0 are at the farthest points (A and A 0 ) with hop distance h from the sink. Let the communication range of a sensor be denoted as r. For ease of explanation, let uv be the Euclidean distance of two points u and v. Obviously, SA ¼ SA0 ¼ hr. • For n = 1: Consider a network shown in the upper right of Fig. 9. The maximum Euclidean distance between two ambiguous sensors is intuitively AA0 . Thus, the maximum hop distance c between such ambiguous sensors 0 0 is 2h because c ¼ AAr ¼ SAþSA ¼ 2hr ¼ 2 h. r r • For n > 1: Consider a network shown in the lower right of Fig. 9. Suppose B and B 0 are two nearest points with hop distance h from the sink. If Assumption 1 holds, AB ¼ A0 B0 ¼ r. The maximum Euclidean distance between two ambiguous sensors with hop distance h is either AA0 or AB0 (A0 B is not considered because A0 B ¼ AB0 ). By the Cosine rule, in triangle DSAB0 ; AB0 can be determined as below. AB02 ¼ SA2 þ SB02  2  SA  SB0  cos a:

ð3Þ

Because SA ¼ hr and SB0 ¼ ðh  1Þr, Eq. (3) can be rewritten as below. 2

AB0 ¼ r2 ðh2 þ ðh  1Þ  2 hðh  1Þ cos aÞ:

ð4Þ

Similarly, by the Cosine rule, in triangle DSAA0 ; AA0 can be determined as below. AA0 ¼ 2r2 h2 ð1  cos aÞ:

ð5Þ

By Eqs. (4) and (5), AA02  AB02 can be represented as below. r2 ð2 hð1  cos aÞ  1Þ: Notably, if AA0 2  AB0 2 > 0, then AA0  AB0 > 0 because AA0 þ AB0 > 0. Recall that a ¼ 2p . We obtain a < p2 2n and further 0 < cosa < 1 if n > 1. Obviously, r2(2h(1  cosa)  1) > 0 because r > 0. Thus AA0 > AB0 is always true if n>1. That is, the maximum Euclidean distancep between two ambiguous sensors is AA0 . As a result, ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi c ¼ h 2ð1  cos aÞ. Consequently, according to cases 1 and 2, Eq. (2) is obtained. h In Fig. 9, let the network size be L · L. Suppose two ambiguous sensors are at the farthest points (A and A 0 ) with hop distance h from the sink. Obviously, L = 2hr. To guarantee that two ambiguous sensors are within the communication range of each other, the maximum hop

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

a

b

1433

c

Fig. 8. Different sink placement scenarios. (a) The sink is at the corner. (b) The sink is at the edge. (c) The sink is within the network. Ndir is assumed to be 8.

Fig. 9. Physical location ambiguity problem. The sink termed S is at the center of the network where N dir ¼ 2p . All sensors in the shadow region have the 2n same hop distance h from the sink.

distance between them should be less that 1. Thus, the relationship among the network size, communication range, and the number of directions is obtained as ( r if n ¼ 1 pffiffi L 6 pffiffiffiffiffiffiffiffiffiffiffiffiffi ð6Þ 2r if n > 1 ð1cos aÞ

is the angle formed by two neighboring axes where a ¼ 2p 2n intersecting at S. Basically, Eq. (6) shows an useful suggestion for a user before sensor deployment, in which the values of L and n are able to be determined according to the requirement of precision in sensor position. For example, L should not exceed 100m for Ndir = 2, 141.4m for Ndir = 4, 261.31m for Ndir = 8, and 512.58m for Ndir = 16, respectively, in case r = 100m. 7. Performance evaluations In this paper, we experiment DLS on different networks by C++. Most sensor localization protocols mention that

the number of sensors and communication range significantly lead to different levels of accuracy. We, thus, conduct numerous simulations to evaluate the influences of these factors on DLS. Much research on localization addresses that correctness of a sensor in physical position is the most important concerns. Thus, in the paper, we also focus on accuracy in directional information by using a metrics called estimated correctness to validate the performance of DLS. Recall that the network model used in DLS involves multiple directional and axis regions. As shown in Fig. 1, the estimate of a sensor locating in R1 is undoubtedly correct if its estimated primary direction code is 00. Note that the VADC system is devised to improve the estimate of the sensor in the axis region. Therefore, although obtaining an ð1Þ ð2Þ incorrect primary direction code, a sensor in A1 or A1 is also correctly estimated in case its estimated auxiliary direction code is 10. Let P be the set of sensors with the accurate estimates in primary direction code. Let A be the set of sensors within the axis region, which obtain the incorrect primary direction code but accurate auxiliary

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

one, corresponding to the axis regions. The sizes of P and A are, respectively, termed NP and NA. Under the consideration of the VDDC system, the correct rate of DLS, termed Cdual is represented as below: C dual ¼

NP þ NA ; Ns

where Ns is the number of sensors in the network. 7.1. Simulation environment In the simulation, the network is a square area with the size of 500m · 500m. Various numbers of directions, such as 4, 8, and 16 are considered. The sink is situated at the center of the network. All sensors are randomly scattered with a uniform distribution within the square area. A probabilistic bound to achieve connected network is addressed in [31]. Therefore, our simulations differ from the numbers of sensors with 500, 600, 700, 800, 900, and 1000. The sink, anchors, as well as sensors have the same communication ranges, which range from 50m to 80m with a step of 10m. Additionally, all simulation results are averaged over 30 simulation runs, respectively. 7.2. Simulation results Extensive simulations are first performed to evaluate the efficiencies of diverse anchor deployment approaches. We, then, show the performances of DLS with the random and greedy methods with respect to different numbers of sensors and communication ranges. Besides, the scenario, where the sink is placed at the edge of the network, is also evaluated. 7.2.1. Effects of anchor deployment strategy Here, we devise two methods, random and greedy methods, to evaluate the impacts of different anchor deployments on sensor localization. The numbers of anchors in both approaches are identical and equal to Ndir. For ease of representation, the random and greedy methods, here, are respectively, indicated by RLS and GLS. In RLS, all anchors are randomly deployed within the whole network, whereas in GLS, whose main idea is to improve the estimated correctness of the critical sensors, anchors are evenly distributed in each direction within the critical regions. Additionally, both of which also use the multi-message decision strategies for direction estimation. Fig. 10 shows the estimated correctness for different anchor deployment strategies. Obviously, GLS outperforms RLS regardless of the number of directions. Due to random deployment of anchors, the average correct rates of RLS for 4, 8, and 16 directions are approximately 30.13%, 35.52%, and 39.82%, respectively. The correct rate actually improves with the increase of Ndir because of the existence of more anchors. In GLS, the average estimated correctness in terms of 4, 8, and 16 directions reach 64.96%, 58.93%, and 43.00%,

100 90 80

Estimated correctness (%)

1434

70 60 50 40 30 20

RLS

10

GLS

DLS

0 4

8

16

Number of directions

Fig. 10. Estimated correctness for different anchor deployment strategies and the number of directions.

respectively. Since GLS exploits the strategy, which places all anchors within the critical region, in sensor localization, more critical sensors enable to obtain the more accurate estimations. Based on the packet propagation, the unknown sensors will learn the high level of accuracy in direction by means of packet propagation. However, the result shows that more directions in GLS obviously incur the disappointing performance. Because all anchors are randomly placed within the critical region of each direction, a critical sensor within multiple anchors’ communication ranges has a high chance to receive the LREQ packets from the anchors with different direction codes. Therefore, a sensor may consequently obtain the erroneous direction. To consider the tradeoff between the number of anchors and the locations of anchors, DLS places all anchors at the communication range from the sink on each axis. Fig. 10 reveals that DLS achieves approximate 94.33%, 86.06%, and 81.09% in correct rates for 4, 8, and 16 directions, respectively. Such anchor deployment technique not only guarantees a critical sensor receives the LREQ packets transmitted from the limited anchors, but also avoids much communication and computation overhead. Therefore, a sensor does not suffer from the interference resulted from the LREQ packets from other directions, and quickly achieves the direction estimation. 7.2.2. Comparison of different localization schemes Fig. 11 shows an example result of spatial sensor distribution, in which the circle and cross, respectively, represent the sensors with the accurate and erroneous estimates. The network with Ndir = 4 consists of 500 sensors, and the communication range of each sensor is 80m. In this case, totally 438 sensors are identified the accurate directions. Namely, DLS achieves an estimated correctness of approximately 88%. Significantly, incorrect estimated sensors always

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

1435

100

500

Estimated correctness (%)

90

y–coordinate (m)

400

300

200

80 70 60 50 40 30 20 10

RLS (Ndir =4)

RLS ( Ndir =8)

GLS (Ndir =4)

GLS (Ndir =8)

RLS ( Ndir =16) GLS (Ndir =16)

DLS (Ndir =4)

DLS ( Ndir =8)

DLS ( Ndir =16)

0

100

50

60

70

80

Communication range (m)

0

0

100

200

300

400

500

Fig. 13. Estimated correctness vs. communication range for different localization schemes.

x–coordinate (m)

Fig. 11. Example of spatial sensor distribution for 500 sensors scattered in the network with Ndir = 4. The blue and red sensors, respectively, represent the sensors with accurate and erroneous estimates.

appear in the vicinity of the axis region due to the receipt of numerous packets from other directions. Figs. 12 and 13, respectively, show the simulation results of the estimated correctness in terms of different numbers of sensors and communication ranges for RLS, GLS and DLS. Among these localization schemes, the performance of DLS is obviously better than those of both GLS and RLS in case of the corresponding Ndir. We conclude that with the aid of our anchor deployment and VDDC strategies, all critical sensors exactly determine their directions, and further enhance the effectiveness of sensor direction learning. Note that the values of Cdual decrease in both DLS and GLS with the increases of Ndir because of a signifi100 90

Estimated correctness (%)

80 70 60 50 40 30 20 10

RLS (Ndir =4)

RLS ( Ndir =8)

GLS (Ndir =4)

GLS (Ndir =8)

GLS (Ndir =16)

DLS (Ndir =4)

DLS ( Ndir =8)

DLS ( Ndir =16)

RLS ( Ndir =16)

0 500

600

700

800

900

1000

Number of sensors

Fig. 12. Estimated correctness vs. number of sensors for different localization schemes. The number of directions in the network is 2n.

cant increase in the size of the axis region. A sensor may receive numerous LREQ packets coming from the sensors placed in the different directions. Such LREQ packets actually enable a sensor to obtain the incorrect estimation although the multi-message decision scheme is exploited. Unlike GLS and DLS, RLS focuses on anchor deployment in the random manner. The numbers of anchors and directions are identical, so the number of anchors increases with the increase of Ndir. The result of RLS significantly reveals that more Ndir leads to the increase of Cdual. Both RLS and GLS, respectively, cause the slight variations in estimated correctness, while DLS results in a steady increase in Cdual if either the number of sensors or the communication range increases. The value of Cdual in RLS is limited between 30% and 40%, and the value of Cdual in GLS is limited between 43% and 65%. The phenomenon in RLS or GLS is resulted from the random anchor deployment, by which unknown sensors within the critical region are likely to obtain the incorrect estimates. Unlike the results of RLS and GLS, the estimated correctness of DLS rises gently for Ndir = 4, 8, and 16 owing to the effectiveness of the proposed anchor deployment strategy. To explicitly outline the effectiveness of DLS, Fig. 14 illustrates the improvements of DLS in correct rate compared to RLS and GLS. Here, the correct rate of each scheme for any number of directions is the average of correct rates generated by all cases with different number of sensors and communication ranges. Significantly, DLS outperforms RLS for Ndir = 4 about 64.20%, for Ndir = 8 about 50.55%, and for Ndir = 16 about 41.27% on average. Besides, DLS also outperforms GLS for Ndir = 4 about 29.37%, for Ndir = 8 about 28.13%, and for Ndir = 16 about 38.09% on average. The observation from Fig. 12 shows that DLS performs best among the three localization approaches for the

1436

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

corresponding number of directions. In DLS, the correct rate slightly increases if the number of sensors increases. Generally, more sensors in the network means that a sensor has more neighbors. Thus, the sensor can receive more LREQ packets, which significantly lead to more accurate estimation owing to the multi-message decision scheme. The curves for Ndir = 4, 8, and 16, in terms of GLS, rise gently with the aid of the multi-message decision mechanism. Although such scheme advantages a sensor to obtain the accurate estimate, some sensors with incorrect direction within the critical region in GLS may incur the poor performance during LREQ packet dissemination. In RLS, all anchors are randomly deployed in the whole network. The estimated correctness has slight variation for different numbers of sensors because of the invalidation of the multi-message decision scheme. In Fig. 13, for DLS, the increase of communication range results in the larger critical region. Thus, the number of critical sensors, whose directions are always accurate also increases. Such critical sensors further lead to the accurate estimate of the unknown sensor within the noncritical region, and, consequently, improve the performance of sensor localization. Additionally, large communication range increases the number of neighbors of a sensor. Therefore, a sensor receives more LREQ packets, and then obtains the accurate direction by means of the multi-message decision scheme. With the characteristic of random anchor deployment, RLS causes slight variation in Cdual for distinct communication ranges. In GLS, with the increase of the communication range, the number of unknown sensors within the critical region may increase owing to more LREQ packets coming from different directions. Such LREQ packets eventually degrade the correct rates of unknown sensors within the critical region. Furthermore, the increase of sensors with erroneous directions may result in the disappointing performance of localization because such incorrect direction information will be propagated via LREQ packets.

Fig. 15 demonstrates the effectiveness of DLS in diverse high density networks. The results are generated by averaging the estimated correctness for various communication ranges in the corresponding number of sensors. Obviously, DLS reaches approximately 96.51%, 90.27%, and 85.59% in accuracy for Ndir = 4, 8, and 16, respectively. The correct rates for Ndir = 4, 8, and 16 all keep rising with large number of sensors because the multi-message decision scheme works significantly. Note that the curves for Ndir = 4, 8, and 16 rise gently after 1000 sensors. We reason that more sensors within the axis regions are likely to deteriorate the correct rate. Overall, the multi-message decision scheme is able to tolerate the estimated errors in spite of large number of sensors within the axis regions. Consequently, DLS can achieve well level of accuracy, especially for high dense WSNs. 7.2.3. Effects of the VADC system As mentioned before, the VADC system is mainly to improve the estimated correctness of the sensor within the axis region. Numerous non-critical applications, such as environmental monitoring can tolerate the inaccuracy in geographic information. Thus, Cdual is obviously unsuitable for such applications to evaluate the estimated correctness. Here, we devise the metrics, Cpri, revised from Cdual to evaluate the effect of the VADC system in sensor localization. Compared to Cdual, Cpri does not consider the effect of NA because NA is related to the VADC system. The Cpri is defined as follows: C pri ¼

NP ; Ns

where Ns is the number of sensors in the network. To demonstrate the effectiveness of the VADC system, we make the simulations about the improvement of the VADC system in terms of different number of sensors and communication ranges, respectively. In Figs. 16 and 17, DLS-VADC means the approach, excluding the VADC system from DLS. Overall, DLS-VADC reaches 82.74%,

100 100

90

90

80

Estimated correctness (%)

Improvement (%)

DLS over RLS

70 60 50 40 30 20 DLS over GLS

10

80 70 60 50 40

DLS (Ndir =4) DLS (Ndir =8) DLS (Ndir =16)

30 20 10

0 4

8

16

Number of directions

0 500

1000

1500

2000

2500

Number of sensors

Fig. 14. Improvements of DLS in correct rates compared to RLS and GLS.

Fig. 15. Performance of scalability of DLS.

3000

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439 100

100

90

90

Estimated correctness (%)

Ndir =4 Ndir =8 Ndir =16

80

Improvement (%)

1437

70 60 50 40 30

80 70 60 50 40

DLS ( Ndir =4) DLS ( Ndir =8) DLS ( Ndir =16)

30

20

20

10

10 0

0 500

600

700

800

900

500

1000

600

700

800

900

1000

Number of sensors

Number of sensors

Fig. 16. Improvements of DLS in correct rates compared to DLS-VADC for different number of sensors. DLS-VADC means the VADC system is not involved in DLS.

Fig. 18. Estimated correctness vs. number of sensors in DLS, where the sink is placed at the edge.

100

100

90 Ndir =4 Ndir =8 Ndir =16

80 Improvement (%)

Estimated correctness (%)

90

70 60 50 40

80 70 60 50 40 DLS ( Ndir =4) DLS ( Ndir =8) DLS ( Ndir =16)

30

30

20

20

10

10

0 50

0 50

60

70

80

Communication range (m)

Fig. 17. Improvements of DLS in correct rates compared to DLS-VADC for various communication ranges. DLS-VADC means the VADC system is not involved in DLS.

67.83%, and 57.05% in the estimated correctness for Ndir = 4, 8, and 16, respectively. As shown in Figs. 16 and 17, the improvement in estimated correctness becomes remarkable with the increase of Ndir. The VADC system significantly leads to the improvement in the estimated correctness for Ndir = 4 about 11.59%, for Ndir = 8 about 18.23%, and for Ndir = 16 about 24.04% on average. Intuitively, the increase of Ndir implies the increase of the sensors within the axis regions. The VADC system enables to increase the value of NA, and consequently enhances the performance in sensor localization. 7.2.4. Edge sink performance Figs. 18 and 19 illustrate the results of DLS used in the network, in which the sink is assumed to be placed at one

60

70

80

Communication range (m)

Fig. 19. Estimated correctness vs. communication range in DLS, where the sink is placed at the edge.

edge. We also evaluate the performance of DLS in a network with 500m · 500m size with regard to the number of sensors and communication range for various Ndir. Figs. 18 and 19 illustrate the results for various number of directions. Overall, either the large number of sensors or the increase of communication range actually advantages sensors to learn the accurate direction. The increase of the value of Ndir dramatically incurs disappointing results due to the enlargement of axis regions. DLS leads to average correct rates for Ndir = 4 about 95.49%, Ndir = 8 about 89.29% and for Ndir = 16 about 84.64%. Actually, placing the sink at the edge of the network causes the network to be partitioned into 2, 4, and 8 distinct regions although Ndir = 4, 8, and 16, respectively. Here, we focus on the correct rates between the corresponding scenarios in the different sink placement methods. The results for Ndir = 4 and 8 shown in both Figs. 12 and

1438

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439

13 outperform the results for Ndir = 8 and 16 revealed in both Figs. 18 and 19, respectively. The reason is that placing the sink at the edge significantly enlarges the axis regions, and further deteriorates the correct rate. 7.2.5. Summary As expected, the estimated correctness increases in case of larger number of sensors because of the marked advantage of the multi-message decision scheme. Additionally, the increase of the communication range is most likely to increase the number of critical sensors aware of the accurate directions. The critical sensors with well-estimated directions then propagate their direction information to the network during the course of packet dissemination. Such behavior, consequently, most probably enables the unknown sensor to obtain the accurate estimate. We find that large Ndir incurs less correct rate owing to the larger percentage of the axis region, which means the sensors within the axis regions most probably obtain the erroneous estimates. However, such scenario is able to be efficiently improved by means of the VADC system. 8. Conclusions To our best knowledge, the paper is the first investigation to the localization based on sensor direction. We introduce an effortless and cost-effective localization solution, DLS, whose main goal is for each sensor to estimate its direction related to the sink. Spatial locality property motivates DLS for direction estimating according to the packets received at a sensor. DLS not only uses the elegant anchor deployment strategy for the improvement of estimated correctness, but also efficiently exploits the virtual dual direction coordinate system for the identification of the precise location of the sensor around the axis. Two important factors, number of sensors and communication range, are studied in the simulations. For the scenario, in which the sink is placed at the center, the average correct rates for Ndir = 4, 8, and 16 are approximately 94%, 86%, and 81%, respectively. In general, DLS outperforms RLS and GLS in direction estimation owing to the use of a well-designed anchor deployment strategy. DLS is wellsuited to high dense WSNs as well. With the aid of the VDDC system, DLS enables a sensor to efficiently recognize that it is near the axis although it obtains an incorrect direction code. Additionally, the proposed DLS is also suitable for the scenario in which the sink is not at the center of the network. For the enhancement of direction determination at each sensor, future studies can investigate techniques, such as the Bayesian inference method, which takes the prior estimates into account, and weight-based approach, which assigns the individual weight to each packet by means of its RSSI value. Furthermore, an efficient routing protocol for WSNs is being designed on the basis of the information of the relative direction and the hop count obtained by DLS.

Acknowledgement This work was supported by the National Science Council of the Republic of China under Grants NSC 95-2221-E032-005 and NSC 95-2218-E-263-001. References [1] J. Ford, Telecommunications with MEMS devices: an overview, The 14th Annual Meeting of the IEEE Lasers and Electro-Optics Society 2 (12) (2001) 415–416. [2] Y.-B. Ko, N.H. Vaidya, Location-aided routing (LAR) in mobile ad hoc networks, Wireless Networks 6 (4) (2000) 307–321. [3] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, A survey on sensor networks, IEEE Communications Magazine 40 (8) (2002) 102–114. [4] H. Yang, B. Sikdar, A protocol for tracking mobile targets using sensor networks, in: Proceedings of the IEEE International Workshop on Sensor Network Protocols and Applications (SNPA), May 2003, pp. 71–81. [5] K. Kalpakis, K. Dasgupta, P. Namjoshi, Efficient algorithms for maximum lifetime data gathering and aggregation in wireless sensor networks, Computer Networks 42 (6) (2003) 697–716. [6] T. Vercauteren, D. Guo, X. Wang, Joint multiple target tracking and classification in collaborative sensor networks, IEEE Journal on Selected Areas in Communications 23 (4) (2005) 714–723. [7] N. Sadagopan, B. Krishnamachari, A. Helmy, The ACQUIRE mechanism for efficient querying in sensor networks, in: Proceedings of the IEEE International Workshop on Sensor Network Protocols and Applications (SNPA), 2003, pp. 149–155. [8] H. Sabbineni, K. Chakrabarty, Location-aided flooding: an energyefficient data dissemination protocol for wireless sensor networks, IEEE Transactions on Computers 54 (1) (2005) 36–46. [9] J.N. Al-Karaki, A.E. Kamal, Routing techniques in wireless sensor networks: a survey, IEEE Wireless Communications 11 (6) (2004) 6– 28. [10] T. Melodia, D. Pompili, I.F. Akyildiz, On the interdependence of distributed topology control and geographical routing in ad hoc and sensor networks, IEEE Journal on Selected Areas in Communications 23 (3) (2005) 520–532. [11] S. Meguerdichian, F. Koushanfar, M. Potkonjak, M.B. Srivastava, Coverage problems in wireless ad hoc sensor networks, in: Proceedings of the IEEE INFOCOM, the Annual Joint Conference of the IEEE Computer and Communications Societies, April 2001, pp. 1380–1387. [12] S. Megerian, F. Koushanfar, M. Potkonjak, M.B. Srivastava, Worst and best-case coverage in sensor networks, IEEE Transactions on Mobile Computing 4 (1) (2005) 84–92. [13] N. Bulusu, J. Heidemann, D. Estrin, GPS-less low cost outdoor localization for very small devices, IEEE Personal Communications Magazine 7 (5) (2000) 28–34. [14] T. He, C. Huang, B.M. Blum, J.A. Stankovic, T. Abdelzaher, Rangefree localization schemes for large scale sensor networks, in: Proceedings of the ACM International Conference on Mobile Computing and Networking (MOBICOM), September 2003, pp. 81–95. [15] Y. Shang, W. Ruml, Y. Zhang, Localization from mere connectivity, in: Proceedings of the ACM International Symposium on Mobile Ad Hoc Networking and Computing (MOBIHOC), June 2003, pp. 201– 212. [16] B.H. Wellenhof, H. Lichtenegger, J. Collins, Global Positioning System: Theory and Practice, Springer-Verlag, Berlin, 1997. [17] A. Savvides, C.-C. Han, M.B. Strivastava, Dynamic fine-grained localization in ad-hoc networks of sensors, in: Proceedings of the ACM International Symposium on Mobile Ad Hoc Networking and Computing (MOBIHOC), August 2001, pp. 8–14.

S.-S. Wang et al. / Computer Communications 30 (2007) 1424–1439 [18] N.B. Priyantha, A. Chakraborty, H. Balakrishnan, The cricket location-support system, in: Proceedings of the ACM International Conference on Mobile Computing and Networking (MOBICOM), August 2000, pp. 32–43. [19] D. Niculescu, B. Nath, Ad hoc positioning system (APS) using AoA, in: Proceedings of the IEEE INFOCOM, the Annual Joint Conference of the IEEE Computer and Communications Societies, March 2003, pp. 1734–1743. ˇ apjun, M. Hamdi, J.-P. Hubaux, GPS-free positioning in mobile [20] S. C ad hoc networks, Cluster Computing 5 (2) (2002) 157–167. [21] P. Bahl, V. Padmanabhan, RADAR: An in-building RF-based user location and tracking system, in: Proceedings of the IEEE INFOCOM, the Annual Joint Conference of the IEEE Computer and Communications Societies, March 2000, pp. 775–784. [22] L. Lazos, R. Poovendran, HiRLoc: High-resolution robust localization for wireless sensor networks, IEEE Journal on Selected Areas in Communications 24 (2) (2006) 233–246. [23] D. Niculescu, B. Nath, DV based positioning in ad hoc networks, Telecommunication Systems 22 (1) (2003) 267–280. [24] X. Ji, H. Zha, Sensor positioning in wireless ad-hoc sensor networks using multidimensional scaling, in: Proceedings of the IEEE INFOCOM, the Annual Joint Conference of the IEEE Computer and Communications Societies, March 2004. [25] S. Vural, E. Ekici, Wave addressing for dense sensor networks, in: Proceedings of the IEEE Workshop on Sensor and Actor Network Protocols and Applications (SANPA), 2004, pp. 56–66. [26] I. Stojmenovic, X. Lin, A loop-free routing for wireless networks, in: Proceedings of the IASTED Conference on Parallel and Distributed Systems, August 1998, pp. 48–55. [27] R. Iyengar, B. Sikdar, Scalable and distributed GPS free positioning for sensor networks, in: Proceedings of the IEEE International Conference on Communications (ICC), May 2003, pp. 338–342. [28] A. Nasipuri, K. Li, A directionality based location discovery scheme for wireless sensor networks, in: Proceedings of the ACM International Workshop on Wireless Sensor Networks and Applications (WSNA), September 2002, pp. 105–111. [29] R. Nagpal, H. Shrobe, J. Bachrach, Organizing a global coordinate system from local information on an ad hoc sensor network, in: Proceedings of the IEEE International Workshop on Information Processing in Sensor Networks (IPSN), April 2003, pp. 48–55. [30] T.S. Rappaport, Wireless Communications - Principles and Practice, Prentice Hall, Englewood Cliffs, NJ, 1996. [31] C. Bettstetter, On the minimum node degree and connectivity of a wireless multihop network, in: Proceedings of the ACM International Symposium on Mobile Ad Hoc Networking and Computing (MOBIHOC), June 2002, pp. 80–91.

Sheng-Shih Wang received the B.S. degree in the department of Computer Science and Information Engineering from Tunghai University, Taiwan, in 1993 and the M.S. degree in the department of Transportation and Communication Management Science from National Cheng Kung University, Taiwan, in 1995, respectively. He received the Ph.D. degree in the department of Computer Science and Information Engineering from Tamkang University, Taiwan, in 2006. In the fall of 2006, he joined the faculty of the department of Information Network Technology at Chihlee Institute of Technology, Taiwan, and served as the Assistant Professor.

1439

From 1999 to 2002, he worked as an R&D Engineer at Siemens Telecommunication Systems Limited, engaged in the design and development of software of switches. Dr. Wang is a member of the IEEE Computer and Communication Societies and Phi Tau Phi Scholastic Honor Society. His current research interests include wireless sensor networks, wireless LANs and MANs, mobile computing, and network protocols design.

Kuei-Ping Shih received the B.S. degree in mathematics from Fu-Jen Catholic University, Taiwan, Republic of China, in June 1991 and the Ph.D. degree in computer science and information engineering from the National Central University, Taiwan, Republic of China, in June 1998. After two years of military obligation, he joined the faculty of the Department of Computer Science and Information Engineering, Tamkang University, Taiwan, Republic of China, as an Assistant Professor in 2000. Currently, he is an associate professor in the Department of Computer Science and Information Engineering, Tamkang University, Taiwan, Republic of China. Dr. Shih served as a guest editor for the International Journal of Wireless and Mobile Computing (IJWMC), special issue on Mobile Systems, E-commerce, and Agent Technology, a Program Area Chair for the IEEE International Conference on Advanced Information Networking and Applications (AINA), 2005, a Technical Track Chair for the IEEE International Conference on Information Technology: Research and Education (ITRE), 2005, a Local Arrangement Co-Chair for International Conference on SCORM (SCORM), 2006, and a Program Committee member for several international conferences. Dr. Shih’s current research interests include wireless LANs and MANs, wireless sensor networks, mobile computing, and network protocols design. Dr. Shih has received the research award of National Science Council, Republic of China, 2000. He is a member of the IEEE Computer Society, the IEEE Communication Society, and the Phi Tau Phi Scholastic Honor Society.

Chih-Yung Chang received the Ph.D. degree in Computer Science and Information Engineering from National Central University, Taiwan, in 1995. He joined the faculty of the Department of Computer and Information Science at Aletheia University, Taiwan, as an Assistant Professor in 1997. He was the Chair of the Department of Computer and Information Science, Aletheia University, from August 2000 to July 2002. He is currently an Associate Professor of Department of Computer Science and Information Engineering at Tamkang University, Taiwan. Dr Chang served as an Associate Guest Editor of Journal of Internet Technology (JIT, 2004), Journal of Mobile Multimedia(JMM, 2005),International Journal of Wireless and Mobile Computing, and a member of Editorial Board of Tamsui Oxford Journal of Mathematical Sciences (2001-2005). He was an Area Chair of IEEE AINA’2005, Vice Chair of IEEE WisCom’2005 and EUC’2005, Track Chair (Learning Technology in Education Track) of IEEE ITRE’2005, Program Co-Chair of MNSAT’2005 and UbiLearn’ 2006, Workshop Co-Chair of INA’2005, MSEAT’2003, MSEAT’2004, and Publication Chair of MSEAT’2005 and SCORM’2006. Dr. Chang is a member of the IEEE Computer Society, Communication Society and IEICE society. His current research interests include wireless sensor networks, mobile learning, Bluetooth radio networks, Ad Hoc wireless networks, and mobile computing.