530 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 15, NO. … Papers/2016 .Net/LSD1624 - A Joint Time...A Joint Time Synchronization and Localization Design for Mobile Underwater Sensor

A Joint Time Synchronization and LocalizationDesign for Mobile Underwater Sensor Networks

Jun Liu,Member, IEEE, Zhaohui Wang,Member, IEEE, Jun-Hong Cui,Member, IEEE,

Shengli Zhou, Fellow, IEEE, and Bo Yang,Member, IEEE

Abstract—Time synchronization and localization are basic services in a sensor network system. Although they often depend on each

other, they are usually tackled independently. In this work, we investigate the time synchronization and localization problems in

underwater sensor networks, where more challenges are introduced because of the unique characteristics of the water environment.

These challenges include long propagation delay and transmission delay, low bandwidth, energy constraint, mobility, etc. We propose

a joint solution for localization and time synchronization, in which the stratification effect of underwater medium is considered, so that

the bias in the range estimates caused by assuming sound waves travel in straight lines in water environments is compensated. By

combining time synchronization and localization, the accuracy of both are improved jointly. Additionally, an advanced tracking algorithm

interactive multiple model (IMM) is adopted to improve the accuracy of localization in the mobile case. Furthermore, by combining both

services, the number of required exchanged messages is significantly reduced, which saves on energy consumption. Simulation

results show that both services are improved and benefit from this scheme.

Index Terms—UWSNs, synchronization, localization, sensor node

Ç

1 INTRODUCTION

IN recent years, underwater sensor networks (UWSNs)have gained significant attention from academic and

industrial researchers due to the potential benefits andunique challenges posed by the water environment.UWSNs have allowed a host of applications to become bothfeasible and effective, including coastal surveillance, envi-ronmental monitoring, undersea exploration, disaster pre-vention and mine reconnaissance. However, due to the highattenuation of radio waves in water, acoustic communica-tion is emerging as the most suitable media. Several charac-teristics specific to underwater acoustic communicationsand networking introduce additional design complexityinto almost every layer of the network protocol stack [1], [2],[3], [4], [5]. For example, low communication bandwidth,long propagation delays, higher error probability, andsensor node mobility are concerns that must be confronted.

Among the services UWSNs can provide, time synchro-nization and localization are very critical, because most

UWSNs applications benefit from or require these two serv-ices. For instance, Time Division Multiple Access (TDMA),one of the commonly used medium access control (MAC)protocols, often requires precise synchronization amongsensor nodes. Additionally, most geographic routing algo-rithms assume the availability of location information [6],[7].

Although localization and synchronization services areclosely related, they are usually studied independently.This is mainly because localization is traditionally studiedfrom the signal processing point of view in radio networks,and synchronization is mainly studied from protocol designpoint of view. However, especially in UWSNs, localizationand synchronization are closely “bonded”. Since the rang-ing is estimated based on time of arrival (TOA) or time dif-ference of arrivals (TDOA) in UWSNs, many localizationalgorithms rely on the time synchronization services. Forexample, in TOA, a popular localization algorithm, synchro-nization is a prerequisite. On the other hand, knowledge oflocation helps time synchronization because it can be usedto estimate propagation delays. Furthermore, both localiza-tion and time synchronization require a sequence of mes-sage exchanges among the nodes. Based on these bondsrelationships, we believe that localization and time synchro-nization could be solved jointly, with two major benefits.First, a joint strategy would save energy, since localizationand synchronization can use only one set of messageexchanges instead of two. This is important for energyconstrained network systems like UWSNs. Second, a jointsolution can help to improve the accuracy of both services.

However, the research on joint design of synchronizationand localization in UWSNs is still limited. Additionally, inUWSNs, all current localization algorithms [8], [9], [10], [11]assume the straight line transmission of acoustic waves. Infact, due to the sound speed variation with depth in the

� J. Liu is with the College of Computing Science and Technology, JilinUniversity, Changchun 130012, China. E-mail: [email protected].

� Z. Wang is with the Department of Electrical and Computer Engineering,Michigan Technological University, 121 EERC Building, 1400 TownsendDrive, Houghton, MI 49931. E-mail: [email protected].

� J.-H. Cui is with the Department of Computer Science and Engineering,University of Connecticut, 371 Fairfield Way U-4155, Storrs, CT 06269.E-mail: [email protected].

� S. Zhou is with the Department of Electrical and Computer Engineering,University of Connecticut, 371 Fairfield Way U-4157, Storrs, CT 06269.E-mail: [email protected].

� B. Yang is with the Department of Electronic Information and ElectricalEngineering, Shanghai Jiao Tong University, 800 Dongchuan Street,Shanghai 200240. E-mail: [email protected].

Manuscript received 7 Jan. 2014; revised 19 Jan. 2015; accepted 16 Feb. 2015.Date of publication 27 Apr. 2015; date of current version 2 Feb. 2016.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference the Digital Object Identifier below.Digital Object Identifier no. 10.1109/TMC.2015.2410777

530 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 15, NO. 3, MARCH 2016

1536-1233� 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

water environment, called “stratification effect”, the realtransmission path usually bends. This will severely affectthe ranging estimation, and in turn affect localizationaccuracy.

In this paper, we propose a joint solution for localizationand synchronization, called JSL, for UWSNs. JSL is a fourphases scheme. For each round of message exchanges, timesynchronization and localization are performed at differentphases. During iterations, the output of synchronization isfed back as the input of localization, and the output of local-ization is fed back as the input of synchronization in thenext round of message exchanges. In this way, synchroniza-tion and localization are interleaved and can benefit eachother by improving the accuracy of both services. Duringthe localization phase, unlike other algorithms that assumesound waves travel in straight lines in the water environ-ment, JSL compensates the stratification effect when per-forming the underwater acoustic ranging, so that thepropagation delay estimation will be significantlyimproved. Furthermore, to account for the mobility charac-teristic in UWSNs, a tracking method—called interactivemultiple model (IMM), is used to predict sensor nodemobility to improve the accuracy of localization.

The rest of this paper is organized as follows. We firstreview the related work summarized in Section 2, and thenintroduce the design challenges in Section 3. We describeJSL in Section 4, and analysis in Section 6. Then simulationresults are presented in Section 7. Finally, we offer our con-clusions and future work in Section 8 and some backgroundknowledge in Appendix A.

2 RELATED WORK

Among all of the time synchronization algorithms for terres-trial wireless sensor networks, RBS [12], TPSN [13], andFTSP [14] are the most popular and representative three.Reference broadcast synchronization (RBS) is a well-knownreceiver-receiver synchronization algorithm. It completelykills errors on sender side. However, RBS requires extramessages to exchange local time-stamps between any twonodes which intend to get synchronized, and its major ideagreatly depends on immediate reception of reference mes-sages. Timing-sync protocol for sensor networks (TPSN) is asender-receiver time synchronization scheme. AlthoughTPSN takes care of propagation delay, it does not take theclock skew into account during synchronization procedure,which severely limits its accuracy. Flooding time synchroni-zation protocol (FTSP) is designed for sniper localization, soit is required to achieve considerably high accuracy. And asFTSP applies flooding technique, it is very robust to the net-work topology change. FTSP is not applicable to UWSNsmainly because it also assumes instant reception. In addi-tion, it requires hardware calibration, which is not a com-plete software solution.

Although there are significantly growing interests inUWSNs in the past several years, the research on timesynchronization for UWSNs is still relatively limited.TSHL [15] is designed for high latency networks, whichcan manage long propagation delays and remain energyefficient. TSHL combines one-way and two-way MAC-layer message delivery. TSHL works well for static

underwater sensor networks, but it cannot handle mobilescenarios as it assumes constant propagation delays amongsensor nodes. MU-Sync [16] is proposed to synchronizenodes in a cluster based UWSNs. Although MU-Sync aimsto solve sensor node mobility issue in UWSNs, it requiresrelatively high message overhead. Furthermore, MU-Syncuses half of the round trip time to calculate one way prop-agation delay, which causes extra errors, especially for fastmoving situations or when ordinary nodes respondto cluster head after experiencing a long time duration.Mobi-Sync [17] is a time synchronization scheme formobile underwater acoustic sensor networks. Mobi-Syncdistinguishes itself by considering spatial correlationamong the mobility patterns of neighboring UWSNs nodes.This enables Mobi-Sync to accurately estimate the longdynamic propagation delays. However, Mobi-Sync onlyworks for dense network as it is based on spatial correla-tion. TSMU [18] is a pairwise synchronization method.It effectively utilizes Doppler effects and incorporates therelative speed of the transmitter and receiver to improvethe dynamic propagation delay estimation. Additionally,Kalman filter and the calibration process are exploited tofurther enhance synchronization accuracy. However,TSMU only show its performance with linear kinematicmodel, which limits its practical applicability.

Traditionally, localization algorithms are either range-based or range-free. In this work, we focus on the range-based localization algorithms. Ranging techniques are eithercommunication-based or connectivity-based. GPS Intelli-gent Buoys (GIBs) [19] and PARADIGM [20] use underwa-ter GPS. Those to-be-localized sensor nodes need tocommunicate with the surface buoys, because of which, thetime synchronization is a base for them in order to converttime information into range information. Hahn and Rice[21] tries to avoid time synchronization by using half of theround-trip time to estimate the propagation delay. Itrequires a sensor node to communicate with multiple sur-face buoys, which may introduce a heavy load of traffic inthe network. A silent positioning scheme is proposed [22]which does not depends on time synchronization. Instead,all the sensor nodes get localized by passively listening tothe beacon messages exchanged among anchor nodes, sothat it requires four noncoplanar anchors that can mutuallyhear each other.

Usually, connectivity-based method is only used whenthere is no direct communication between nodes and sen-sors, where range estimation is achieved based on networkconnectivity. Niculescu and Nath [23] proves that euclideanmethod performs the best in anisotropic topologies, butwith more cost on computation and communication. In [8],it proves that with short communication range amonganchor nodes, Euclidean method can be adopted for 3Dunderwater localization. Zhang and Cheng [24] relaxes thislimitation by proving it is also working when the anchornodes communication range is long. However, both [8] and[24] are suffering from heavy traffic due to using floodingmechanism. Mirza and Schurgers [25] introduced “SLMP”,which applies some prediction schemes to localization algo-rithm to reduce its overhead. In SLMP, anchor nodes con-duct linear prediction by taking advantages of the inherenttemporal correlation of underwater object mobility pattern.

LIU ET AL.: A JOINT TIME SYNCHRONIZATION AND LOCALIZATION DESIGN FOR MOBILE UNDERWATER SENSOR NETWORKS 531

While each ordinary sensor node predicts its location by uti-lizing the spatial correlation of underwater object mobilitypattern, weighted-averaging its received mobilities fromother nodes. However, because the prediction is based ontemporal and spatial correlations, the algorithm only worksin dense network.

Tian et al. proposed the first localization and synchroni-zation scheme for 3D underwater acoustic networks in [26],using atomic multilateration and iterative multilaterationtechniques. In this scheme, the to-be localized sensor nodeobtains time and localization information from the anchornodes, which are localized on the surface of the water inorder to perform time synchronization and localization. Thedrawbacks of this algorithm is that it ignores the clock skewwhen doing time synchronization, which will lead to fre-quent re-synchronization. Additionally, the time values andlocations are calculated in the same formulas as variablesand they do not improve upon each other, which makes this“joint” solution weak.

3 DESIGN CHALLENGES

In order to achieve effective time synchronization and local-ization in UWSNs, four critical challenges have to beaddressed.

3.1 Stratification Effects

Underwater acoustic localization usually relies on TOAmeasurements, which are converted into range estimates.However, in some scenarios (e.g., deep water environ-ments), the water medium is inhomogeneous and the soundspeed varies depending on several parameters, such as tem-perature, pressure and salinity. As a result, sound waves donot necessarily travel in straight lines. Ignoring this stratifi-cation effect could lead to considerable bias in the rangeestimates.

3.2 Long Propagation Delays

Any software based time synchronization approaches usingmessage exchanges have to face several uncertainties whichcould affect accuracy. Those uncertainties include sendingtime, accessing time, transmission time, propagation time,reception time, interrupt handling, encoding time, decodingtime and byte alignment time [27], [28]. In UWSNs, amongthese uncertainties, the propagation time is dominant dueto the low propagation speed of acoustic signals. Such long

propagation latencies heavily affect the accuracy of timesynchronization algorithms which assume instant synchro-nization message reception [15], [16]. Therefore, in order toachieve more accurate time synchronization in UWSNs,estimating and compensating the long propagation delay isa must do job. Besides of propagation delay, the transmis-sion delay also has to be considered. It is comparable topropagation delay due to the low bandwidth of acousticchannel, especially when the packet size is big and the dis-tance between sensor nodes are short.

3.3 Sensor Node Mobility

While terrestrial sensor networks are usually static, sensornodes in an underwater environment often have passivemobility caused by water currents or proactive mobilitycoming with mobile platforms, which makes localizationchallenging. This is because in such a situation, it is difficult,if not impossible, to estimate the real time distance betweentwo sensor nodes, which will in turn affect localizationaccuracy.

The mobility also complicates time synchronization bycausing continuous changes of propagation delays. How-ever, most of the existing time synchronization schemes usehalf of the round trip time to calculate one way propagationdelay. Due to node mobility, the propagation delays on theway to and from nodes are not necessarily identical, espe-cially when nodes move at a high speed. Therefore, toimprove the time synchronization accuracy, the sensornode mobility should be considered.

3.4 Energy Constraints

Underwater sensor nodes are usually powered by batteries,for which it is not easy to replenish. Therefore, the life timeof a sensor node is restricted by the limited power supply.For this reason, synchronization and localization overheadneeds to be carefully controlled. Synchronization or locali-zation protocols requiring frequent message exchanges arenot suitable in UWSNs. Therefore, an effective synchroniza-tion and localization algorithm should be developed with alimited message exchange overhead.

4 DESCRIPTION OF JSL

For this work, the network architecture consists of threetypes of sensor nodes as shown in Figs. 1 and 2. Surfacebuoys equipped with GPS serve as “satellite nodes”.

Fig. 1. Network scenario 1. Fig. 2. Network scenario 2.


Anchor nodes are powerful sensor nodes which cancommunicate with a certain number of surface buoysdirectly to become synchronized and localized. Both surfacebuoys and anchor nodes act as reference nodes and ordi-nary nodes are those sensor nodes or autonomous under-water vehicles (AUVs) which intend to be synchronizedand localized.

When the water medium is inhomogeneous in terms ofpressure, salinity and temperature, the range estimate is notlinearly proportional to the TOA measurement in UWSNs,as illustrated in Figs. 3 and 4 [29]. In order to avoid the biasfrom assuming the straight line transmission, in JSL, weadopt an improved range estimation scheme which com-pensates the above stratification effect. A detailed descrip-tion of the ranging procedure can be found in Appendix A.Note that, we consider localization and synchronizationbetween two network nodes. Although the underwater sen-sor network is three-dimensional, any two network nodesdefines a plane which is perpendicular to the sea bottom.Therefore, it suffices to represent the position of any net-work node by two-dimensional coordinates on the plane.

4.1 Overview

The procedure of JSL consists of four major phases, DataCollection, Synchronization, Localization and Iteration, asshown in Fig. 5. Those four phases are executed once within

each round of message exchange and iteration phase is thebridge to the procedure in next round of messageexchanges. In Phase I, an ordinary sensor node acquires ref-erence time and location information from neighboringreference nodes. In Phase II, the synchronization process isperformed by ordinary nodes based on the informationobtained in Phase I. Phase II consists of four steps. First, inthe first round of message exchanges, a node’s rough posi-tion is estimated by using the TDOA method. After the firstround, the rough position is the estimate of localizationprocedure in last round of message exchanges. After thestratification effect is compensated, the propagation delay iscalculated. Next, JSL performs linear regression to synchro-nize the ordinary sensor node by using all the time stampsit collected and all the propagation delays it calculated in allprevious rounds of message exchanges, which is followedby the update of the corresponding propagation delays forthis round of message exchanges. During Phase III, JSL car-ries out the localization process based on the estimatedpropagation delays in Phase II. Additionally, a trackingalgorithm, IMM, is used to improve localization accuracy,such that the final location estimates are combined from theestimates based on Fermat’s principle and the predictedvalue from IMM. Phase IV is an iteration process, the posi-tion estimated in Phase III acts as input to Phase II to replacethe rough position in next round of message exchange. Thewhole procedure completes until there are no more messageexchanges.

4.2 Phase I: Data Collection and Rough PositionEstimation

As shown in Fig. 6, an ordinary node initiates the procedureby broadcasting a Req message to its neighboring referencenodes. The ordinary node stamps the sending time T1

obtained at the MAC layer, right before the message leaves.Upon receiving the Req message, before the reception anddecoding of the message, each reference node immediatelymarks its local time as t2. Then similarly, after a time

Fig. 3. Sound velocity profile.

Fig. 4. Sound wave propagation paths.

Fig. 5. The work flow of JSL.


interval tr (waiting for the hardware sending-receiving tran-sition and avoiding collisions), each of the reference nodesends back a Res message containing its location informa-tion, t2, and sending time t3. When receiving the Res mes-sage, the ordinary node marks its receiving time T4. Fig. 6shows the message exchange process among the ordinarynode and reference nodes. Fig. 7 demonstrates the proce-dure between each pair.

After receiving Res messages from all the referencenodes, if it is the first round of message exchanges, the ordi-nary node will estimate its rough position with TDOA. Wecall it “rough position” because of two reasons. First, at thismoment, the ordinary node is not synchronized yet, there-fore the obtained TDOA is not precise. Second, withoutconsidering stratification effect, straight line transmissionsare still assumed in this phase.

In JSL, ordinary nodes have clocks aiming to becomesynchronized with the clocks of reference nodes. Hence wehave:

T ¼ u � tþ b; (1)

where T stands for the measured time of the ordinary node;t is the reference time; u and b are the relative clock skewand offset, respectively.

At this stage, in order to calculate ordinary node’s roughposition, we first assume the ordinary node has been syn-chronized. This means we assign an initial clock skew “1”,and an initial clock offset “0”. We take reference node 1 asthe base node and therefore comparing with reference node“1”, the time difference for reference node “n” is:

Dtn1 ¼ tn � t1; n ¼ 2; . . . ; N: (2)

where t1 denotes the estimate of the TOA for base reference

node “1”, d1 stands for the corresponding distance for t1, tndenotes the estimate of the TOA for base reference node

“n”, dn stands for the corresponding distance for tn, and

Dtn1 stands for the estimate of the TDOA for base referencenode “n”. Therefore, the distance difference dn1 = dn - d1 canbe estimated as:

dn1 ¼ hDtn1; (3)

where h denotes the sound average propagation speed.

Denote (xn, yn, zn) as the position of reference node “n”,and (xr, yr, zr) as the position of the node to be localized.Based on dn ¼ dn1 þ d1, we have [30]

ðdn1 þ d1Þ2 ¼ x2n þ y2n þ z2n � 2xnxr � 2ynyr � 2znzr þ d21;

which can be simplified as

xnxr þ ynyr þ znzr ¼ 1

2x2n þ y2n þ z2n � d2n1

� �� dn1d1: (4)

We assume that each sensor node is equipped with a depthsensor, which leads to

xnxr þ ynyr ¼ 1

2x2n þ y2n þ z2n � d2n1

� �� dn1d1 � znzr: (5)

Define

M ¼x2 y2x3 y3

..

. ...

xN yN

0BBB@

1CCCA;D ¼

�d21�d31...

�dN1

0BBB@

1CCCA;U ¼

�z2�z3...

�zN

0BBB@

1CCCA;

Q ¼ 1

2

x22 þ y22 þ z22 � d221

x23 þ y23 þ z23 � d231

..

.

x2N þ y2N þ z2N � d2N1

0BBB@

1CCCA;P ¼ xr

yr

� �:

The input-output relationship in (5) can be recast as

MP ¼ Qþ d1Dþ zrU: (6)

The least squares (LS) estimate of the ordinary node’sposition is obtained as

P ¼ MyQþ d1MyDþ zrM

yU; (7)

where y stands for the pseudo-inverse of matrix.If it is not the first round of message exchange, the rough

position will be the estimate of last round localizationprocedure.

4.3 Phase II: Synchronization

The synchronization phase consists of three major steps:propagation delay estimation, linear regression and propa-gation delay update, which are explained in the following.

Fig. 7. Message exchange for each pair.

Fig. 6. Message exchanges.


4.3.1 Step I: Propagation Delay Estimation

According to Section A, in order to increase the accuracy ofpropagation delay estimation when the water medium isinhomogeneous, it is necessary to deal with the bias associ-ated with the straight-line propagation assumption (Pleaserefer to Appendix A).1 As shown in Fig. 8, since we knowthe ordinary node’s rough position, represented with P 0

r inFig. 8, a rough rr � rs represented with rr

0 � rs is easy to becalculated, where rr and rs are the projections of the ordi-nary node’s position and the reference node’s position ontothe y-axis, respectively. Therefore, with Eq. (27), by per-forming a numerical search,2 a constant “C” can be fixed. InEq. (26), by knowing “C”, the propagation delay “T” can beestimated with stratification effect compensated. Using thismethod, the ordinary node calculates the correspondingpropagation delay “T” for each exchanged message. Thus,in the end, the ordinary node will collect a set of propaga-tion delays that consist of t1 and t2.

4.3.2 Step II: Linear Regression

During this step, the ordinary node updates its estimates ofclock skew u and offset b using the ordinary least square esti-mation (OLSE) method based on the collected time stampsand the calculated propagation delays in all the previousrounds of message exchanges. In this work, we assume theexchanged messages have the same packet size, bandwidthand spectral efficiency, so that the transmission delay is fixedand known to all the messages, which is estimated as tr=ð packetsize

bandwidth�spectralefficiencyÞ. Corresponding to the ith round of

message exchange, we denote T1½i� and T4½i� as the timestamps collected by the ordinary node, t2½i� and t3½i� as thetime stamps collected by a reference node, and t1½i� and t2½i�as propagation delays in the two-way message transmission.At the Mth round of message exchange, we denote theupdated estimates of clock skew and offset as u½M� and b½M�,respectively. Based on Eq. (1), the input-output relationshipfor clock skew and offset estimation is cast as

T1½i� ¼ u½M� � t1½i� þ b½M� þ e1½i�T4½i� ¼ u½M� � t4½i� þ b½M� þ e4½i�t2½i� ¼ t1½i� þ t1½i� þ tr þ e2½i�t3½i� ¼ t4½i� � t2½i� � tr þ e3½i�

8>><>>: (8)

for i ¼ 1; . . . ;M, where e1½i� and e4½i� are the time stampmeasurement noise at the ordinary node, and e2½i� and e3½i�are the time stampmeasurement noise at the reference node.

We define

B ¼

T1½1�T1½2�...

T1½M�

0BBBBB@

1CCCCCAor

T4½1�T4½2�...

T4½M�

0BBBBB@

1CCCCCA;

A ¼

t2½1� � t1½1� � tr 1

t2½2� � t1½2� � tr 1

..

. ...

t2½M� � t1½M� � tr 1

0BBBBB@

1CCCCCAor

t3½1� þ t2½1� þ tr 1

t3½2� þ t2½2� þ tr 1

..

. ...

t3½M� þ t2½M� þ tr 1

0BBBBB@

1CCCCCA;

C ¼ u½M�b½M�

!:

Using the least-squares method, the new clock skew u½M�and offset b½M� are estimated as

C ¼ ðATAÞ�1ATB: (9)

4.3.3 Step III: Update Propagation Delay

When the synchronization process completes, the ordinarysensor node becomes synchronized, and a new draft clockskew u and offset b are estimated. The propagation delay t1and t2 can then be updated. Since the synchronization proce-dure is an optimization process, the propagation delays tendto bemore accurate than the estimated ones in Step I.

4.4 Phase III: Localization

The localization procedure is actually the opposite processof estimating propagation delay (please refer to AppendixA). After Phase II, using Eq. (26), with updated propagationdelay “T”, the constant “C” can be fixed by doing numericalsearch. Then in Eq. (27), the rr (rr � rs) can be estimated.

Since �r-�s is known with depth sensor, so “R” in Fig. 8 is:

R ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðrr � rsÞ2 þ ð�r � �sÞ2

q: (10)

With this method, the ordinary node computes “R” for

each reference node to build a vector consisting of “Rs”:

ðR1; R2; R3; . . . ; RMÞT :

Considering TDOA, we can observe that “R” is just the“d” and therefore it is easy to convert the above vector to“D” in Eq. (7). Next, with updated “D”, the position “P”can be estimated with Eq. (7). Using this method, for eachrun of message exchanges, one position will be estimated,such that the ordinary node will have a sequence of updatedposition measurements.

Fig. 8. Illustration of the stratification effect.

1. In typical shallow water environments without sound speed strat-ification, the stratification effect compensation discussed in the sequelcan be omitted.

2. In this work, the bisection method, also called binary searchmethod is applied to do the searching.


However, in the mobile scenario where either ordinarynodes and reference nodes are moving, it is hard and nearlyimpossible to estimate the real time distance or propagationdelay. Note that our localization method is based on the esti-mates of the distances. Therefore, some errors might be intro-duced. In order to reduce the effect frommobility, we adopt atracking algorithm to predict sensor nodes’ position at nextmeasurement time. In this work, localization estimates comefrom a combination of prediction andmeasurements.

However, given the potential of multiple moving modesof each node, a universal state equation is not possible. Wehence employ an IMM estimator to accommodate the multi-ple maneuver patterns. A brief introduction on the IMMestimator is provided in the sequel, and a detailed descrip-tion can be found in the book [31].

IMM is an adaptive estimation approach. Different frommany other methods which assume a particular moving pat-tern of the node, the IMM filter incorporates all the possiblemoving patterns of the node, by running a bank of filters inparallel with each filter corresponding to one particularmoving pattern. And the overall state estimate is a certaincombination of these model-conditional estimates. In thepresence of unpredictable mobility patterns of the networknode, the model mismatch error can be modeled as the pro-cess noise in the state equations of the bank of filters [31].

In this paper, with an assumption of Gaussian noise, weconsider two moving patterns of the node:

� uniform motion. Moving along the straight line with aconstant velocity, which can be modeled by a Kal-man filter.

� maneuver. Coordinated turn with a constant turn rateand a constant speed, which can be modeled by anextended Kalman filter.

As shown in Fig. 9, we have the two filters running in par-allel, and the combination of the estimates from the two fil-ters yields the final estimated position of the node.3 Thus, in

Fig. 9, Mk stands for measured position at time k, which is

calculated from Eq. (7), Exk is the predicted position using

the paralleled filters, andNk is the final estimated position.Given that nodes’ moving pattern can change from

one mode to another, the IMM estimate of the node loca-tion is computed as a weighted summation of the esti-mates from multiple mode-matching filters. A Markovchain transition matrix characterizing the transition prob-ability of the node’s moving pattern from one mode toanother, is introduced to update the weighting coeffi-cients. Except that the initial condition of each filter is acombination of the model-conditioned estimates of thetwo filters at the preceding state, operation of each filteris identical to that of the Kalman filter or the extendedKalman filter, depending on whether the state equationunder the considered mode is linear or not.

4.5 Phase IV: Iteration

The iteration process is the bridge to the whole procedure inthe next round of message exchanges. During iteration, thepositions estimated in the localization phase are input asrough position for the procedure in next round of messageexchanges. Besides, all the time stamps and correspondingpropagation delays are input to synchronization phase innext round for reference. The whole JSL procedure com-pletes until no more message exchanges.

5 JSL IN MULTI-HOP

As illustrated in Fig. 10, in a multi-hop network which ismore realistic, usually, not every ordinary node can reachthree anchor nodes. In this case, to achieve network widelocalization and synchronization, a multi-hop scheme isdemanded. This section intends to show how JSL isextended from one-hop network to multi-hop and solve net-work wide localization and synchronization.

Imagine that to achieve multi-hop or even network widesynchronization and localization, when anchor nodes cannot cover all the ordinary nodes which intend to get syn-chronized and localized, more other reference nodes arerequired. In fact, after those ordinary nodes which arereachable from three anchor nodes get localized and syn-chronized, they can be the potential candidates. In this case,the coverage area which is reachable from more than threereference nodes increases recursively as nodes are newly

Fig. 9. Block diagram of the tracking routine.

Fig. 10. Multi-hop network scenario.

3. For a system with non-Gaussian noise, it has been shown that the(extended) Kalman filter suffers significant performance degradation[32]. Hence, in the presence of non-Gaussian noise and with the noisedistribution parameters available, advanced filtering algorithms, suchas the bootstrap filter [32] and the particle filter [33], can be used.


synchronized and localized join the reference node set.However, considering the network wide localization andsynchronization error, the reference nodes need to bepicked selectively. Eligible reference nodes have to meet theaccuracy requirement first, and then they can assist otherordinary nodes to get synchronized and localized.

For network wide synchronization and localization, wepropose a hierarchical approach, dividing the ordinary nodeslocalization and synchronization process into two sub-pro-cesses. At first, ordinary nodes reachable from at least threeanchor nodes get localized and synchronized with JSL, fol-lowing the steps introduced in previous sections. And foreach synchronized and localized ordinary node, it is associ-atedwith a confidence value, representedwith “h”:

h ¼ nd�Pni¼1½ðT1 � t1Þ þ ðT2 � t2Þ�i

’nTl: (11)

In Eq. (11), the T1 and T2 are propagation delays estimatedin Eq. (26), and the subscript “ 1 ”,“ 2 ” stand for the two cor-responding delays for in each message exchange. “ i ” is theindex for the message exchange, and “ n ” is the total num-ber of message exchange. “ d ” is the threshold for the toler-ant propagation delay difference estimated in localizationand synchronization process, which is used to represent thesynchronization and localization accuracy with JSL. “ ’ ” isthe coefficient for “ T l ”, which counts the time elapse afterthe sensor node get synchronized and localized, so that theconfidence value drops with time process after the referencenode gets localized and synchronized. If an ordinary node iseligible to be a reference node after it is localized andsynchronized is based on the confidence value:

h � 0 eligible,h < 0 not eligible.

�

At second step, ordinary node not reachable to three anchornodes will try to get synchronized and localized with thehelp of the new reference nodes. And some of them maybecome new reference nodes to assist more localization andsynchronization process. Thus, JSL is extended to be anetwork wide solution through the association with a recur-sive strategy. Fig. 11 describes the process for each ordinarynode to get synchronized and localized in the network.

6 CONVERGENCE ANALYSIS

Considering the whole procedure of JSL shown in Fig. 5,data collection is to collect sample data, say “time stamps”.The first TDOA is to launch the JSL procedure by providinginitial positions. The propagation delay is the bridge toconnect time synchronization and localization. With themethod introduced in Section A, the propagation delaywith stratification effect compensated is estimated based onrough position (first run) or the updated position from pre-vious localization routine. With this propagation delay, theordinary node becomes synchronized by applying linearregression (with OLSE). The propagation delay is input tolocalization procedure to estimate the distance, after locali-zation, the position is used to estimate the propagation forthe next round of message exchange.

This section intends to prove the iteration process willconverge. Considering the iteration procedure in Fig. 5, toprove it converges, we need to in turn prove that more accu-rate propagation delay estimation can make synchroniza-tion more precise, more accurate synchronization processcan obtain more accurate propagation delay, and moreaccurate propagation delay can improve the localizationaccuracy.

Considering the iteration procedure, at the end of theðn� 1Þthmessage exchange, we have4:

b½n� 1� ¼ 1

n� 1

Xn�1

i¼1

ðT ½i� � t½i�Þ þ dn�1

hþ tr: (12)

At nthmessage exchange, we have,

dnh

¼ 1

n

Xni¼1

ðt½i� � T ½i�Þ þ b½n� 1� � tr: (13)

By substituting Eq. (12) in Eq. (13), we have:

dnh

¼ 1

n

Xni¼1

ðt½i� � T ½i�Þ � 1

n� 1

Xn�1

i¼1

ðt½i� � T ½i�Þ þ dn�1

h: (14)

It can be verified that:

ðdn � dn�1Þ2h2

¼ 1

nðn� 1ÞXn�1

i¼1

ðt½i� � T ½i�Þ þ 1

nðt½i� � T ½i�Þ

!2

;

(15)

which shows that ðdn � dn�1Þ2 ! 0 as n ! 1. Hence, thelocalization result converges as the number of messagesincreases. Meanwhile, (12) reveals that synchronizationresult converges along with the localization result.

Fig. 11. Ordinary node localization and synchronization process.

4. Note that, we use Req message as the ðn� 1Þth message. If it isResmessage, T ½i� and t½i� need to be switched in the equations.


7 PERFORMANCE EVALUATION

7.1 Simulation Settings

The simulations are implemented in Aqua-Sim, an NS-2based simulator specifically designed for underwater sensornetworks [34]. It can effectively simulate acoustic signalattenuation, packet collisions, multipath, fading and so onin underwater sensor networks. In this work, we applybroadcast MAC and static routing in MAC layer andnetwork layer. 10 underwater sensor nodes are randomlydistributed in a 2;000m� 2;000m� 2;000m region. And thetransmission range of sensor node are 4;000 meters whichmakes it an one-hop network. The sound velocity profile(SVP) is the same as shown in Fig. 3. For JSL without stratifi-cation, the sound speed is constant as 1;500meters/second.

We define node density as the expected number of nodesin a nodes neighborhood, hence node density is equivalentto node degree. We control the node density by changingthe communication range of each node while keeping thearea of deployment the same. Range (i.e., distance) meas-urements between nodes are assumed to follow normal dis-tributions, with real distances as mean values and standarddeviations to be one percent of real distances. The error onthe reference position from the anchor follows a normal dis-tribution with 0:1 as the mean value and 0:01 as the stan-dard derivation. Without loss of generality, the inherentskew of the ordinary node is predefined randomly with the

range from 104 to 105 ppm, and it remains unchangedduring the time synchronization process. The clock offset ofthe ordinary node is initialized randomly from 20ms to 2 s.In addition, the response time tr is fixed as 1 s. Each refer-ence message is 40 bytes, and the bit rate is 6:2 kbps, whichcorresponds to the real OFDM modem developed in ourUWSN lab [35]. In terms of the water currents, we considerthe kinematic model in [36], which is commonly used inunderwater network research. We also evaluate the impactof water currents by using two other models, namely ran-dom walk and a meandering current mobility model, inSection 7.2.7. The moving speed of the kinematic model canbe described as follows:

Vx ¼ k1�v sin ðkk2xÞ cos ðkk3yÞ þ k1� cos ð2kk1tÞ þ k4;

Vy ¼ ��v cos ðkk2xÞ sin ðkk3yÞ þ k5;

(

(16)

where Vx represents the speed at x-axis, Vy stands for thespeed at y-axis. k1; k2; k3; �; v are variables which are closelyrelated to environment factors such as tides and bathyme-try. These parameters change in different environments.k4; k5 are random variables which are used to simulate therandom factors. In our simulation, we assume k1; k2 to berandom variables which are normally distributed with amean p and a standard deviation 0:1p. k3 follow a normaldistribution with 2p as the mean value and the standardderivation 0:2p. � is also normal distributed with 3 as themean value and 0:3 as the standard derivation. v alsofollows a normal distribution with 1 as the mean value and0:1 as the standard derivation. k4; k5 are random variableswhich are subject to normal distribution with 1 as mean val-ues and 0:1 as standard derivations. k is a coefficientcontrolling the frequency of changing direction.

In this simulation, we assume that sensor nodes followtwo mobility patterns: uniform motion and maneuver withcoordinated turn. Therefore, an IMM-tracker with the abovetwo modes is used.5 Mode 1 is used for the uniform motionwith a constant velocity, while mode 2 is used for themaneuver with coordinated turn. The two modes aredescribed by a Kalman filter and an extended Kalman filter,respectively. The transition probabilities of the two modesused in simulations are

½pij� ¼ 0:95 0:050:05 0:95

� i; j ¼ 1; 2: (17)

7.2 Results and Analysis

7.2.1 Performance On Localization Accuracy

Fig. 12 shows localization accuracy of four algorithmsincluding JSL, JSL without stratification compensation,SLMP II) and TDOA which are pure localization algorithmsacting as benchmarks. The normalized MSE (mean squareerror) stands for NMSE ¼ E½ðr�r

r Þ2�, where

r ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðx� xÞ2 þ ðy� yÞ2

q. SLMP is a localization algorithm

which adopts linear prediction schemes to reduce its over-head, and to make fair comparison, we only count the locali-zation process with anchor nodes as reference nodes.Results disclose that JSL performs much better than theother three, and that is because it does not assume thestraight line acoustic transmission. Additionally, couplingtime synchronization and localization together, JSL makesthem help each other to improve the accuracy of both. JSLwithout stratification compensation is worse than SLMPwhich indicates the importance of the stratification effect,but it is still better than TDOA since it benefits from the jointdesign. From Fig. 12 we can also observe that, with morereference sensor nodes, the results become better, and thatis because when more reference nodes are involved, moredata can be collected in the rough position estimations andsynchronization procedure. Both of these will help toimprove localization accuracy.

Fig. 12. Localization error versus Number of reference nodes.

5. In the real scenario, sensor nodes can have various mobility pat-terns. The IMM-tracker can be extended to have multiple mode-match-ing filters to address these different patterns.


7.2.2 Performance On Synchronization Accuracy

Fig. 13 illustrates how error grows after time synchronizationcompletion with different algorithms. The x-axis is Tc � Tst,where Tc is sample time point after the synchronization andTst is when the time synchronization stops. And y-axis is thetime drift. e.g. JSL, JSL-without stratification compensation,MU-Sync and TSHL. As discussed above, as time passes, theskew causes increasing errors. Therefore, this comparisonactually demonstrates different accuracies on synchroniza-tion that these three algorithms can achieve. Fig. 13 comparesthe performance of the three algorithms using an identicalmessage overhead. The result shows that JSL performs betterthanMU-Sync and TSHL, and that is because of two reasons.First, JSL considers the range bias caused by the stratificationeffect, which will finally improve the accuracy of both locali-zation and synchronization. The second is the benefit of thejoint design, which is also why JSL-without stratificationcompensation ismore precise thanMU-Sync and TSHL.

7.2.3 Convergence Analysis

Figs. 14 and 15 demonstrate how the synchronization andlocalization performance converges with the number of iter-ations. From the results we can easily observe that withmore iterations, both synchronization and localizationerrors decrease. This is because in JSL, more iterations

means that more messages are involved, more sample datawill be collected, so that the estimated skew and offset willbe more precise. This will help to improve the propagationdelay estimation, which will in turn improve localizationaccuracy. Since during the iteration process, the next roundtime synchronization is using the position informationproduced by localization in the current round to improveits accuracy. Hence, in JSL localization and synchronizationare helping each other to improve the accuracy of both.

7.2.4 Impact of Stratification Effect

Figs. 14 and 15 also show that with the same amount of itera-tions, JSL with stratification effect compensation can achievehigher accuracy than thatwithout stratification compensationin both time synchronization and localization. This is becauseignoring the stratification effect would lead to a considerablebias in the range estimates. This bias will be delivered tothe propagation delay estimation, and since propagationdelay is the key to connect time synchronization and localiza-tion, the error caused by the bias will eventually affect theaccuracy of both time synchronization and localization.

7.2.5 Performance of IMM

Fig. 16 shows the contribution of the tracking mechanism tolocalization accuracy. From the result, we can clearly see that

Fig. 13. Synchronization accuracy.

Fig. 14. Convergence on Synchronization versus number of iterations.

Fig. 15. Convergence on Localization versus number of iterations.

Fig. 16. Effect of tracking.


with the help of IMM tracking, localization error is alwayslower than without using any tracking algorithm. Althoughin our simulation setting, as the coefficient “k” increases, thesensor node’s velocity changes too fast in direction, the track-ingmechanism still helps.

7.2.6 Sample Rate for Synchronization Process

Table 1 shows the different sample rates of the four algo-rithms. The sample rate represents the efficiency of theexchanged messages to linear regression. In the four algo-rithms, TSHL and JSL have the highest sample rate. With-out considering dynamic propagation delays, TSHL andJSL employ linear regression over single direction commu-nication during the skew estimation process. As shown insynchronization phase both A and B has two sets of values,which means every exchanged message except the messageused to calculate offset can be taken as a sample data pointduring linear regression. Therefore, their message samplerate is close to 100 percent. MU-Sync has the lowest mes-sage sample rate, as it adopts two-way message exchanges,and assumes the propagation delays on the way back andforth are identical. Accordingly, only one way or half ofthose exchanged messages can be utilized as sample data.In this regard, the messages sample of MU-Sync is as lowas 50 percent. In Mobi-Sync, although super nodes respondwith two messages for one request message, both of thetwo response messages can be applied as sample points.This is because the two messages are estimated respec-tively, so its sample rate is 2=3.

7.2.7 Effect of Mobility Pattern

Figs. 17 and 18 evaluate the impact of mobility pattern ofunderwater sensor nodes on the localization and synchroni-zation. To compare with the pattern we selected in this work,

which is a kinematic model widely used in underwater sen-sor networks, we run the simulation on another well knownmobility pattern-random walk, and a meandering CurrentMobilityModel introduced in [37] which has velocity as:

$ðx; y; tÞ ¼ �tanh y�BðtÞsinðkðx�ctÞÞ½1þk2B2ðtÞcos2ðkðx�ctÞÞ�1=2

� ; (18)

where BðtÞ ¼ Aþ e � cos ðvtÞ. In JSL, when take the recom-mended setting for this mobility model in [37] as follows:A ¼ 1:2; c ¼ 0:12; k ¼ 2=7:5;v ¼ 0:4; e ¼ 0:3. Fig. 17 showsthe impact on localization accuracy and Fig. 18 demonstratesthe impact on synchronization accuracy. From the result, wecan conclude that, the mobility pattern has slight influence onthe localization and synchronization accuracy, because it willaffect the efficiency of tracking algorithm. Since for randomwalk, the velocity at last time point has nothing to dowith thevelocity at the current point, any tracking scheme will losetheir efficacy, and that explains why the results based onrandom walk are worse for both localization and synchroni-zation. And meandering Current Mobility Model is morecomplicated than kinematic model and that is why its accu-racy is a little bit worse than kinematic.

7.2.8 Localization and Synchronization Coverage

Fig. 19 shows that the localization and synchronizationcoverage increases with both the node density and thepercentage of anchor nodes in the whole network. Thisis reasonable because when the node density increases,the connectivity of the network increases, more ordinarynodes can be covered by anchor nodes, and there will bemore new reference nodes. In turn, more ordinary nodescan be localized and synchronized. Similarly, when moresensor nodes are anchor nodes in the whole network,there will be more ordinary nodes that are covered bythe anchor nodes both directly and indirectly, whichmakes easier for the ordinary node to be synchronizedand localized, leading to an increased localization andsynchronization coverage.

7.2.9 Network Wide Localization Error

Fig. 20 demonstrates the network wide localization error,which is the average localization error of all the ordinaryFig. 17. Effect on localization.

TABLE 1Sample Rate for Synchronization Process

TSHL 1:0

JSL 1:0MU � Sync 0:5Mobi� Sync 0:67

Fig. 18. Effect on synchronization.


nodes which are localized with JSL. It shows that thelocalization error decreases with the increase of node den-sity and anchor node percentage. This is because whennode density increases, anchor nodes’ coverage increases,so that more ordinary nodes are directly localized byanchor nodes. Consider that the localization error willaccumulate when the localized ordinary nodes performas reference nodes to localize other ordinary nodes. Asmore ordinary nodes are localized by anchor nodes, thenetwork wide localization accuracy increases. Similarly,with a greater percentage of anchor nodes, more ordinarynodes are localized by anchor nodes, so that the errordecreases.

7.2.10 Network Wide Synchronization Error

Fig. 21 demonstrates the network wide synchronizationerror, which is the average synchronization error of all theordinary nodes which are synchronized with JSL. Similar tothe localization process, it shows that the synchronizationerror decreases with the increase of node density andanchor node percentage. This is because when more ordi-nary nodes are directly synchronized by anchor nodes, thesynchronization accuracy will increase, which is based onthat the synchronization error will accumulate when thesynchronized ordinary nodes’ time are taken as inputs tosynchronize other ordinary nodes. Similarly, with a greater

percentage of anchor nodes, more ordinary nodes aresynchronized by anchor nodes, so that the error decreases.

8 CONCLUSIONS AND FUTURE WORK

In this paper, we presented JSL, a joint solution for timesynchronization and localization in UWSN. It compen-sates the stratification effect in the underwater environ-ment instead of assuming straight line transmission.Furthermore, synchronization and localization are closelycoupled and help each other to improve the accuracy ofeach other. An advanced tracking algorithm IMM isadopted to further improve accuracy. Our simulationresults show that JSL can achieve high accuracy for bothsynchronization and localization.

In the future, we plan to implement and evaluate the JSLin real underwater environment, since the device for testingthe SVP is getting less expensive. And we would alsoconsider how a time varying clock skew and time-variabletransmission rate affect the performance of JSL.

APPENDIX ARANGING WITH STRATIFICATION EFFECT

COMPENSATION

In this paper, we consider an isogradient sound velocityprofile in which the sound speed is only depth-dependent,as illustrated in Fig. 8. This assumption applies well to thedeep-water environments where the temperature and salin-ity are almost constant, and the sound speed is mainlydetermined by the water pressure [38].

Based on the assumption that the sound velocity profileis only depth dependent, it is sufficient to consider that boththe sender and receiver are on a two-dimensional plane, asshown in Fig. 8. We therefore describe the stratificationeffect compensation procedure in a two-dimensional coor-dinate system.

Let vð�Þ denote the SVP as a function of depth �, anddefine ðrs; �sÞ and ðrr; �rÞ as the position of the sender andthe receiver, respectively. We further assume that i) the SVPvð�Þ is available, which can be measured prior to experi-ments or obtained empirically, ii) the position of the senderðrs; �sÞ is known as a priori, iii) the depth of the receiver can

be roughly estimated as �r through a depth-sensor. OnceFig. 20. Network wide localization error.

Fig. 21. Network wide synchronization error.Fig. 19. Localization and synchronization coverage.


the estimation of rr is obtained, the range estimate can becalculated as

R ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðrr � rsÞ2 þ ð�r � �sÞ2

q: (19)

Next, we focus on the estimation of rr based on the noisy

TOA measurement T , the SVP vð�Þ, the sender depth �s and

the receiver depth �r.It is known that the travel time for an acoustic ray along a

possible path S is expressed as

T ¼ZS

1

vðsÞ ds: (20)

Denote r ¼ fð�Þ as the function of the acoustic ray repre-sented by cartesian coordinates in the plane expanded by r-and �-axes, and define its first-order derivative with respectto � as f 0ð�Þ :¼ dr=d�. We have

ds ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffid�2 þ dr2

p¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1þ f 0ð�Þ2

qd�: (21)

Notice that vðsÞ is only depth-dependent. The travel timecan be equivalently written as

T ¼Z xir

�s


qvð�Þ d�: (22)

According to Fermat’s principle, the true travel path isthe one which minimizes the travel time, so that fð�Þ can beobtained via

d

d�

@

@f 0


qvð�Þ ¼ 0; (23)

which leads to

f 0ð�Þvð�Þ


q ¼ C; (24)

where C is an integration constant. The travel path fð�Þ canbe then obtained based on

f 0ð�Þ ¼ Cvð�Þffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1� ½Cvð�Þ�2

q ; (25)

where the constant C can be determined based on the ToA

measurement T via

T ¼Z �r

�s

1

vð�Þ1ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

1� ½Cvð�Þ�2q d�: (26)

The rr of the receiver is therefore

rr ¼ fð�rÞ ¼Z �r

�s

f 0ð�Þdz ¼Z �r

�s

Cvð�Þffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1� ½Cvð�Þ�2

q d�: (27)

It has been proved in [29] that estimates yr and R arebias-free and reach the Cramer-Rao lower bound.

ACKNOWLEDGMENTS

Jun Liu is the corresponding author. J.-H. Cui and S. Zhouhave an ownership interest in AquaSeNT, which is a marinesensor and communication technology firm that developsunderwater wireless communication technology. This workis supported by the CHINA National Science FoundationNSFC61401438.

REFERENCES

[1] I. F. Akyildiz, D. Pompili, and T. Melodia, “Underwater acousticsensor networks: Research challenges,” Ad Hoc Netw., vol. 3, no. 3,pp. 257–279, Mar. 2005.

[2] J.-H. Cui, J. Kong, M. Gerla, and S. Zhou, “Challenges: Buildingscalable mobile underwater wireless sensor networks for aquaticapplications,” IEEE Netw., Special Issue Wireless Sensor Netw.,vol. 20, no. 3, pp. 12–18, May/Jun. 2006.

[3] J. Heidemann, Y. Li, A. Syed, J. Wills, and W. Ye, “Research chal-lenges and applications for underwater sensor networking,” inProc. IEEE Wireless Commun. Netw. Conf., 2006, pp. 228–235.

[4] J. Partan, J. Kurose, and B. N. Levine. (2006, Sep.) A Survey of Prac-tical Issues in Underwater Networks,” in Proc. ACM Int. WorkshopUnderWater Netw., pp. 17–24 [Online]. Available: http://prisms.cs.umass.edu/brian/pubs/partan.wuwnet2006.pdf

[5] Z. Zhou, S. Le, and J.-H. Cui, “An OFDM based MAC protocol forunderwater acoustic networks,” in Proc. ACM Int. WorkshopUnderWater Netw., Sep. 2010, pp. 6:1–6:8.

[6] P. Xie, L. Lao, and J.-H. Cui, “VBF: Vector-based forwarding pro-tocol for underwater sensor networks,” in Proc. 5th Int. IFIP-TC6Conf. Netw. Technol., Services, Protocols; Perform. Comput. Commun.Netw.; Mobile Wireless Commun. Syst., May 2006, pp. 1216–1221.

[7] H. Yan, Z. Shi, and J.-H. Cui, “DBR: Depth-based routing forunderwater sensor networks,” in Proc. 7th Int. IFIP-TC6 Netw.Conf. AdHoc Sensor Netw., Wireless Netw., Next Generation Internet,May 2008, pp. 72–86.

[8] Z. Zhou, J.-H. Cui, and S. Zhou, “Localization for large-scaleunderwater sensor networks,” in Proc. 6th Int. IFIP-TC6 Conf. AdHoc Sensor Netw., Wireless Netw., Next Generation Internet, Atlanta,GA, USA, May 2007, pp. 108–119.

[9] Z. Zhou, J.-H. Cui, and A. Bagtzoglou, “Scalable localization withmobility prediction for underwater sensor networks,” presentedat the IEEE INFOCOM, Mini-Conf., Phoenix, AZ, USA, 2008.

[10] Kenneth and D. Frampton, “Acoustic self-localization in a distrib-uted sensor network,” IEEE Sensors J., vol. 6, no. 1, pp. 166–172,Feb. 2006.

[11] N. Bodhi, H. Balakrishnan, E. Demaine, and S. Teller, “Mobile-assisted localization in wireless sensor networks,” in Proc. IEEEINFOCOM, Mar. 2005, pp. 172–183.

[12] J. Elson, L. Girod, and D. Estrin, “Fine-grained network timesynchronization using reference broadcasts,” in Proc. 5th Symp.Oper. Syst. Design Implementation, Boston, MA, USA, Dec. 2002,pp. 147–163.

[13] S. Ganeriwal, R. Kumar, and M. B. Srivastava,, “Timing-sync pro-tocol for sensor networks,” in Proc. 1st Int. ACM Conf. EmbeddedNetw. Sensor Syst., 2003, pp. 138–149.

[14] M. Maroti, B. Kusy, G. Simon, and A. Ledeczi, “The flooding timesynchronization protocol,” in Proc. 2nd Int. ACM Conf. EmbeddedNetw. Sensor Syst., 2004, pp. 39–49.

[15] A. Syed and J. Heidemann, “Time synchronization for highlatency acoustic networks,” in Proc. IEEE Infocom, Barcelona,Spain, 2006, pp. 1–12.

[16] N. Chirdchoo, W.-S. Soh, and K. C. Chua, “MU-Sync: A time syn-chronization protocol for underwater mobile networks,” pre-sented at the Int. Workshop UnderWater Networks, SanFrancisco, CA, USA, Sep. 15, 2008.

[17] J. Liu, Z. Zhou, Z. Peng, and J.-H. Cui, “Mobi-sync: Efficient timesynchronization for mobile underwater sensor networks,” in Proc.IEEE Globecom, Dec. 2010.

[18] J. Liu, Z. Wang, Z. Peng, M. Zuba, J.-H. Cui, and S. Zhou, “TSMU:A time synchronization scheme for mobile underwater sensornetworks,” in Proc. IEEE Globecom Conf., Dec. 2011, pp. 1–5.


[19] C. Bechaz and H. Thomas, “GIB system: The underwater GPSsolution,” in Proc. Eur. Conf. Underwater Acoust., May 2001.

[20] T. C. Austin, R. P. Stokey, and K. M. Sharp, “PARADIGM: Abuoy-based system for AUV navigation and tracking,” in Proc.IEEE Conf. Exhib. Oceans, 2000, pp. 935–938.

[21] M. Hahn and J. Rice, “Undersea navigation via a distributedacoustic communication network,” in Proc. Turkish Int. Conf.Acoust., Istanbul, Turkey, Jul. 2005.

[22] X. Cheng, H. Shu, Q. Liang, and D.-C. Du, “Silent positioning inunderwater acoustic sensor networks,” IEEE Trans. Veh. Technol.,vol. 57, no. 3, pp. 1756–1766, May 2008.

[23] D. Niculescu and B. Nath, “Ad hoc positioning system (APS),” inProc. IEEE Global Telecommn., 2001, vol. 5, pp. 2926–2931.

[24] Y. Zhang and L. Cheng, “A distributed protocol for multi-hopunderwater robot positioning,” in Proc. IEEE Proc. Int. Conf. Robot.Biomimetics, 2004, pp. 480–484.

[25] D. Mirza and C. Schurgers, “Motion-aware self-localization forunderwater networks,” in Proc. 3rd ACM Int. Workshop WirelessNetw. Testbeds, Exp. Eval. Characterization, 2008, vol. 57, no. 3,pp. 51–58.

[26] C. Tian, W. Liu, J. Jin, and Y. Wang, “Localization and synchroni-zation for 3D underwater acoustic sensor networks,” in Proc. 4thInt. Conf. Ubiquitous Intell. Comput., 2007, pp. 622–631.

[27] H. Kopetz and W. Schwabl, “Global time in distributed real-timesystem,” Technische Universitat Wien, Wien, Austria, Tech. Rep.15/89, 1989.

[28] H. Kopetz and W. Ochsenreiter, “Clock synchronization in dis-tributed real-time systems,” IEEE Trans. Comput., vol. C-36, no. 8,pp. 933–939, Aug. 1987.

[29] C. R. Berger, S. Zhou, P. Willett, and L. Liu, “Stratification effectcompensation for improved underwater acoustic ranging,” IEEETrans. Signal Process., vol. 56, no. 8, pp. 539–546, Aug. 2008.

[30] P. Carroll, S. Zhou, H. Zhou, X. Xu, J.-H. Cui, and P. Willett,“Underwater localization and tracking of physical systems,”J. Electr. Comput. Eng., vol. 2012, p. 11, 2012, Doi:10.1155/2012/683919.

[31] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan, Estimation with Appli-cations to Tracking and Navigation. New York, NY, USA: Wiley-Interscience, 2001.

[32] N. Gordon, D. Salmond, and A. Smith, “Novel approach to non-linear/non-Gaussian Bayesian state estimation,” Radar Signal Pro-cess., IEE Proc., vol. 140, no. 2, pp. 107–113, Apr. 1993.

[33] M. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “Atutorial on particle filters for online nonlinear/non-GaussianBayesian tracking,” vol. 50, no. 2, pp. 174–188, Feb. 2002.

[34] P. Xie, Z. Zhou, Z. Peng,H. Yan, T. Hu, J.-H. Cui, Z. Shi, Y. Fei, and S.Zhou, “Aqua-Sim: An NS-2 based simulator for underwater sensornetworks,” in Proc. MTS/IEEE Biloxi - Marine Technol. Our Future:Global Local Challenges, 2009, pp. 1–7.

[35] H. Yan, L. Wan, S. Zhou, Z. Shi, J.-H. Cui, J. Huang, and H. Zhou,“DSP based receiver implementation for OFDM acousticmodems,” Elsevier J. Phys. Commun., vol. 5, pp. 22–32, 2012.

[36] A. C. Bagtzoglou and N. A, “Chaotic behavior and pollution dis-persion characteristics in engineered tidal embayments: A numer-ical investigation,” J. Am. Water Res. Assoc., vol. 43, pp. 207–219,2007.

[37] C. Antonio, F. Paparella, L. F. Vieira, M. Erol, and M. Gerla, “Themeandering current mobility model and its impact on underwatermobile sensor network,” in Proc. 27th Conf. Comput. Commun.IEEE, INFOCOM, 2008, pp. 221–225.

[38] M. Stojanovic and J. Preisig, “Underwater acoustic communica-tion channels: Propagation models and statistical character-ization,” vol. 47, no. 1, pp. 84–89, Jan. 2009.

Jun Liu (S’10-M’13) received a BS degree in 2002, fromWuhan Univer-sity, an MS degree in 2011 and a PhD degree from the University ofConnecticut (UCONN), Storrs, all in Computer Science and Engineer-ing. His major research interests include time synchronization, localiza-tion, deployment for underwater acoustic networks, and also interestedin operating system, cross layer design. Dr. Liu served as technicalreviewers for many premier journals and conferences. He is a memberof the IEEE.

Zhaohui Wang (S’10-M’13) received a B.S. degree in 2006, from Bei-jing University of Chemical Technology (BUCT), an MS degree in 2009,from the Institute of Acoustics, Chinese Academy of Sciences (IACAS),Beijing, China, and a PhD degree from the University of Connecticut(UCONN), Storrs, all in electrical engineering. Dr. Wang has been withthe Department of Electrical and Computer Engineering at the MichiganTechnological University (Michigan Tech), Houghton, as an assistantprofessor since 2013. Her research interests lie in the areas of wirelesscommunications and statistical signal processing, with recent focuson the multicarrier modulation and signal processing techniquesfor wireless communications and networking in underwater acousticenvironments. Dr. Wang served as technical reviewers for many pre-mier journals and conferences. She was recognized as an OutstandingReviewer by the IEEE Journal of Oceanic Engineering in 2012 and2014, respectively. In 2013, she was honored with the Women of Inno-vation Award by the Connecticut Technology Council.

Jun-Hong Cui received the BS degree in computer science from JilinUniversity, China, in 1995, the MS degree in computer engineering fromthe Chinese Academy of Sciences in 1998, and the PhD degree in com-puter science from the University of California, Los Angeles, in 2003.Currently, she is on the faculty of the Computer Science and Engineer-ing Department at the University of Connecticut, Storrs. Her researchinterests include the design, modeling, and performance evaluation ofnetworks and distributed systems. Recently, her research mainlyfocuses on exploiting the spatial properties in the modeling of networktopology, network mobility, and group membership, scalable and effi-cient communication support in overlay and peer-to-peer networks, andalgorithm and protocol design in underwater sensor networks. She isactively involved in the community as an organizer, a TPC member, anda reviewer for many conferences and journals. She is a guest editor forACM Mobile Computing and Communications Review and Elsevier AdHoc Networks. She cofounded the first ACM International Workshop onUnderWater Networks (WUWNet 2006) and now serves as the WUW-Net steering committee chair. She is a member of the IEEE, ACM, ACMSIGCOMM, ACM SIGMOBILE, IEEE Computer Society, and IEEECommunications Society.

Shengli Zhou received the BS and MSc degrees in electrical engineer-ing and information science from the University of Science and Technol-ogy of China, Hefei, in 1995 and 1998, respectively. He received thePhD degree in electrical engineering from the University of Minnesota,Minneapolis, in 2002. He is currently the Charles H. Knapp associateprofessor in electrical engineering in the Department of Electrical andComputer Engineering, University of Connecticut, Storrs. His generalresearch interests include wireless communications and signal process-ing. His recent focus has been on underwater acoustic communicationsand networking. He served as an associate editor for the IEEE Transac-tions Wireless Communications from 2005 to 2007 and the IEEE Trans-actions Signal Processing from 2008 to 2010. He is now an associateeditor for the IEEE Journal of Oceanic Engineering. He received the2007 ONR Young Investigator award and the 2007 Presidential EarlyCareer Award for Scientists and Engineers. He held a United Technolo-gies Corporation Associate Professorship in Engineering Innovationfrom 2008 to 2011 and has been a member of the Connecticut Academyof Science and Engineering since 2010. He is a member of the IEEE.

Bo Yang obtained his PhD in Electrical Engineering from the City Uni-versity of Hong Kong in 2009. Prior to joining Shanghai Jiao Tong Uni-versity in 2010, he was a postdoctoral researcher at the Royal Instituteof Technology (KTH), Sweden, from 2009 to 2010 and a visiting scholarat the Polytechnic Institute of New York University in 2007. He is now aprofessor in Shanghai Jiao Tong University. He is also a cyber scholarwith the Cyber Joint Innovation Center founded by Zhejiang University,Tsinghua University, and Shanghai Jiao Tong University. His currentresearch interest includes game theoretical analysis and nonlinear opti-mization of communication networks and smart grid. He is on the edito-rial board of Digital Signal Processing-Elsevier and in the TPC ofseveral international conferences. Dr. Yang has been the principle/co-investigator in several research projects funded by NSF of China, Swed-ish Governmental Agency for Innovation Systems, and the US air force.He is a Member of IEEE and ACM.

" For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


Documents

530 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 15, NO. … Papers/2016 .Net/LSD1624 - A Joint Time...A Joint Time Synchronization and Localization Design for Mobile Underwater Sensor