detection, synchronization, and doppler scale estimation with

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. XX, NO. X, SEP. 2008 1

Detection, Synchronization, and Doppler ScaleEstimation with Multicarrier Waveforms in

Underwater Acoustic CommunicationSean F. Mason, Christian R. Berger,Student Member, IEEE,Shengli Zhou,Member, IEEE, and

Peter Willett,Fellow, IEEE

Abstract—In this paper, we propose a novel method fordetection, synchronization and Doppler scale estimation forunderwater acoustic communication using orthogonal frequencydivision multiplex (OFDM) waveforms. This new method involvestransmitting two identical OFDM symbols together with a cyclicprefix, while the receiver uses a bank of parallel self-correlators.Each correlator is matched to a different Doppler scaling factorwith respect to the waveform dilation or compression. Wecharacterize the receiver operating characteristic in terms ofprobability of false alarm and probability of detection. Wealso analyze the impact of Doppler scale estimation accuracyon the data transmission performance. These analytical resultsprovide guidelines for the selection of the detection thresholdand Doppler scale resolution. In addition to computer-basedsimulations, we have tested the proposed method with realdata from an experiment at Buzzards Bay, MA, Dec. 15, 2006.Using only one preamble, the proposed method achieves similarperformance on the Doppler scale estimation and the bit errorrate as an existing method that uses two linearly-frequency-modulated (LFM) waveforms, one as a preamble and the otheras a postamble, around each data burst transmission. Comparedwith the LFM based method, the proposed method works with aconstant detection threshold independent of the noise level and issuited to handle the presence of dense multipath channels. Moreimportantly, the proposed approach does not need to buffer thewhole data packet before data demodulation, which facilitatesfuture development of online realtime receiver for multicarrierunderwater acoustic communications.

Index Terms—Underwater acoustic communication, OFDM,wideband channel, detection, synchronization and Dopplerscaleestimation.

I. I NTRODUCTION

Various data transmission schemes are being actively pur-sued for underwater acoustic (UWA) communications, includ-ing multicarrier modulation in the form of orthogonal fre-quency division multiplexing (OFDM) [1]–[4], single carriertransmission with time-domain sparse-channel equalization[5] or frequency-domain equalization [6], and multi-input

Manuscript received February 29, 2008, revised July 13, 2008, acceptedSeptember 7, 2008. This work is supported by the ONR YIP grantN00014-07-1-0805, the NSF grant ECS 0725562, the NSF grant CNS 0721834, andthe ONR grant N00014-07-1-0055. This work was partially presented atthe IEEE/MTS OCEANS conference, Kobe, Japan, April 2008. The editorcoordinating the review of this manuscript and approving itfor publicationwas Dr. J. Preisig.

The authors are with the Department of Electrical and Computer Engineer-ing, University of Connecticut, 371 Fairfield Way U-2157, Storrs, Connecticut06269, USA (email:{seanm, crberger, shengli, willett}@engr.uconn.edu).

Digital Object Identifier 00.0000/JSAC.2008.000000

multi-output (MIMO) techniques combined with single carrier[7], [8] or multicarrier [9] transmissions. These transmissionschemes are often examined viaoffline data processing basedon recorded experimental data. Towards the development ofanonlineunderwater acoustic receiver, detection and synchro-nization are important, yet often overlooked tasks.

Typically, synchronization entails a known preamble, whichis easily detected by the receiver, being transmitted priortothe data. Existing preambles used in underwater telemetryare almost exclusively based on linearly frequency modulated(LFM) signals, also known as Chirp signals [10]. This is due tothe fact that LFM signals have a desirable ambiguity functionin both time and frequency, which matches well to the un-derwater channel, which is characterized by its large Dopplerspread. However, the receiver algorithms are usually matched-filter based, which try to synchronize a known template to thesignal coming from one strong path, while suppressing otherinterfering paths. This approach suffers from the followingtwo deficiencies: first, the noise level at the receiver has tobe constantly estimated to achieve a constant false alarm rate(CFAR), usually accomplished using order statistics; second,its performance will degrade in the presence of dense andunknown multipath channels.

Due to the slow propagation speed of acoustic waves, thecompression or dilation effect on the time domain waveformneeds to be considered explicitly. Once a Doppler scaleestimate is obtained, a resampling procedure is usually appliedbefore data demodulation [11]. One method to estimate theDoppler scale is to use an LFM preamble and an LFMpostamble around each data burst [11], so that the receiver canestimate the change of the waveform duration. This methodestimates the average Doppler scale for the whole data burst.As such it requires the whole data burst to be buffered beforedata demodulation, which prevents online real-time receiverprocessing.

In this paper, we propose the use of multicarrier wave-forms as preamble for underwater acoustic communications.A preamble that consists of two identical OFDM symbolspreceded by a cyclic prefix (CP) is used. This training patternhas been studied extensively in wireless OFDM systems forradio channels, see e.g., [12], [13], and has been included aspart of the training preamble in the IEEE 802.11a/g standards[14]. The receiver effectively correlates the received signalwith a delayed version of itself, since, thanks to the CP

0000–0000/00$00.00c© 2008 IEEE

2 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. XX, NO. X, SEP. 2008

structure, the repetition pattern persists even in the presence ofunknown multipath channels [12]. However, the synchroniza-tion algorithms that work in wireless radio channels will notperform well in dynamic underwater acoustic channels due tothe large waveform expansion or compression, which changesthe repetition period to some unknown value.

We combine techniques used in underwater telemetry withthe elegant, non-parametric self-correlator approach used inradio channels to perform detection, synchronization andDoppler scale estimation based on multicarrier waveforms.Weuse a bank of parallel branches, with each branch using a self-correlator matched to a different repetition period. Detectionof a data transmission burst is declared when any of thebranches leads to a correlation metric larger than a thresholdvalue. The branch with the largest metric also yields theDoppler scale estimate and coarse synchronization point. Wecharacterize the receiver operating characteristic (ROC)bydeveloping an analytical expression for the probability offalse alarm and a Gaussian approximation for the probabilityof detection. We also analyze the impact of Doppler scaleestimation accuracy on the data demodulation performance.These analytical results provide guidelines for selectingthedetection threshold and Doppler scale resolution. Comparedwith the LFM-preamble based approach, the proposed methodhas the following advantages: (1) the detection threshold isbetween 0 and 1, and doesn’t depend on the channel oroperating SNR; (2) it has a very good detection performance,which is based on the signal energy fromall paths ratherthan only a single path; (3) it leads to accurate Doppler scaleestimation; (4) after coarse timing and resampling, it allowsthe use of fine timing algorithms developed for radio channels,such as [15], [16]; (5) it has a low-complexity implementationwhen the self-correlation metric is computed recursively [17].

In addition to computer-based simulations, we have testedthe proposed method with real data from an experiment atBuzzards Bay, MA, Dec. 15, 2006. Using only one preamble,the proposed method achieves similar performance on theDoppler scale estimation accuracy and the bit error rate asthose methods presented in [4], which are based on theLFM preamble and postamble. However, the proposed methodneeds neither a postamble nor to buffer the whole data burstbefore demodulation. This facilitates future developmentof anonline realtime receiver for multicarrier underwater acousticcommunications.

Note that the proposed method provides apoint estimateofthe Doppler scale at the beginning of a data burst. Therefore,the proposed method is best suited for application scenarioswhere the Doppler scale stays constant or varies slowly duringthe data burst duration. When the Doppler scale changes fast,the point estimates need to be frequently updated, e.g., byshortening of the data burst, or, through the use of a trackingalgorithm based on point estimates available via insertingmultiple synchronization blocks into a long data burst atregular intervals.

The rest of this paper is as follows. The system model isdescribed in Section II, and the proposed receiver algorithm ispresented in Section III. Detection performance is determinedin Section IV, and analysis of the impact of Doppler scale

CP x x guard zeros

ZP OFDM ZP OFDM

preamble data transmission

Fig. 1. A preamble, consisting of two identical OFDM symbolsand a cyclicprefix (CP), precedes the data transmission which uses zero padding.

mismatch on the data demodulation performance is investi-gated in Section V. Section VI contains numerical results, bothfrom simulation and from real data. Section VII contains theconclusion.

II. SYSTEM MODEL

To avoid power consumption in the guard interval betweenOFDM symbols, zero-padded (ZP) OFDM is preferred fordata transmission [3], [4]. For synchronization purposes,thepreamble consists of two identical OFDM symbols togetherwith a cyclic prefix. The overall transmission structure isshown in Fig. 1. The OFDM parameters can be selectedindependently for the preamble and the data transmissions.

Suppose thatK0 subcarriers have been used in the preamble,and one OFDM symbol is of durationT0. The subcarrierspacing is then1/T0 and the bandwidth isB = K0/T0.Let fc denote the carrier frequency, andfk = fc + k/T0

denote the frequency for thekth subcarrier in passband, wherek ∈ S = {−K0/2, . . . ,K0/2 − 1}. Let Tcp denote the lengthof the CP, and define a rectangular window of lengthTcp+2T0

as

q(t) =

{

1 t ∈ [−Tcp, 2T0],

0 otherwise.(1)

The preamble in baseband can be written as

x(t) =∑

k∈S

s[k]ej2π kT0

tq(t) (2)

and the corresponding passband signal is

x(t) = Re

{

ej2πfct∑

k∈S

s[k]ej2π kT0

tq(t)

}

= Re

{

∑

k∈S

s[k]ej2πfktq(t)

}

, (3)

wheres[k] is the transmitted symbol on thekth subcarrier.The channel impulse response for a time-varying multipath

underwater acoustic channel can be described by

c(τ, t) =∑

p

Ap(t)δ (τ − τp(t)) , (4)

where Ap(t) is the path amplitude andτp(t) is the time-varying path delay. As in [4], we adopt the following assump-tions for algorithm development1.

1In short, our algorithms are developed based on assumptionsA1) and A2).Hence, the application of the proposed method shall be limited to underwaterenvironments where A1) and A2) hold true approximately. Methods suitedfor application scenarios where different paths have quitedifferent Dopplerscales warrant further investigation.

MASON et al.: DETECTION, SYNCHRONIZATION, AND DOPPLER SCALE ESTIMATION WITH MULTICARRIER WAVEFORMS IN UAC 3

A1) All paths have a similar Doppler scaling factora suchthat

τp(t) ≈ τp − at. (5)

In general, different paths could have different Doppler scalingfactors, e.g., see field test results in [18], [19]. The methodproposed in this paper is based on the assumption that allthe paths have approximately the same Doppler scaling factor.When this is not the case, part of useful signals are treated asadditive noise, which could increase the overall noise varianceconsiderably [4]. However, we find that as long as the domi-nant Doppler shift is caused by the direct transmitter/receivermotion, as it is the case in our experiment, this assumptionseems to be justified. The value ofa is usually less than0.01 when the relative speed between the transmitter and thereceiver is below 10 m/s.A2) The path delaysτp, the gainsAp, and the Doppler scaling

factor a are constant over the preamble duration2T0 +Tcp.

The preamble duration is usually around 50 ms to 200 ms.Assumption A2) is reasonable within such a duration, as thechannel coherence time is usually on the order of seconds.

When the passband signal in (3) goes through the channeldescribed in (4) and (5), we receive:

y(t) = Re

{

∑

k∈S

s[k]ej2πfk(1+a)t

×∑

p

Apq(

(1 + a)t− τp)

e−j2πfkτp

}

+ n(t), (6)

where n(t) is the additive noise. Defineτmax = maxp τp,which is usually less than the CP lengthTcp. Using thedefinition of q(t) in (1), we obtain

y(t) = Re

{

∑

k∈S

Hks[k]ej2πfk(1+a)t

}

+ n(t),

t ∈ Tcyclic =

[

−Tcp − τmax

1 + a,

2T0

1 + a

]

, (7)

where we define the channel transfer function

C(f) =∑

p

Ape−j2πfτp (8)

and the frequency response on thekth subcarrier as

Hk = C(fk). (9)

Converting the passband signaly(t) to baseband, such thaty(t) = Re

{

y(t)ej2πfct}

, we have:

y(t) =∑

k∈S

Hks[k]ej2π

(

kT0

+afk

)

t+ n(t)

= ej2πafct∑

k∈S

Hks[k]ej2π k

T0(1+a)t + n(t), (10)

for t ∈ Tcyclic andn(t) is the noise at baseband.As expected for CP-OFDM, we observe a cyclic convo-

lution between the signal and the channel in the specifiedinterval, where each subcarrier is only multiplied by the

windowlength Nl

windowlength N1

windowlength NL

|x|

|x|

|x|

max(x)conversion to

baseband

parallel self-correlators

≥ Γth

Fig. 2. To account for the time compression/dilation, multiple parallelbranches are used, each tuned to a certain periodNl.

corresponding frequency response. Due to the wideband natureof the underwater channel, the frequency of each subcarrierat baseband has been shifted differently by an amount ofafk = afc + ak/T0.

III. T HE PROPOSEDALGORITHM

The transmitter sends a baseband waveform embedding arepetition pattern as

x(t) = x(t+ T0), −Tcp ≤ t ≤ T0. (11)

Such a repetition pattern persists in the received signaly(t)even after time-varying multipath propagation as

y(t) = e−j2π a1+a

fcT0y

(

t+T0

1 + a

)

,

− Tcp − τmax

1 + a≤ t ≤ T0

1 + a, (12)

as can be verified from (10). However, the receiver knowsneither theperiod nor the waveform due to the unknownmultipath channel. The problem is then to detect a patternlike (12) from the incoming signal, and infer the repetitionperiod to find the Doppler scale.

Although the problem of estimating an unknown delay(synchronization) and Doppler scale has been studied in theliterature, [20]–[22], these approaches assume a non-dispersivechannel and are therefore based on matched filter processingand the ambiguity function. Ref. [23] has treated randomechos, but the suggested approach relies on estimating theunknown channel transfer function jointly with delay andDoppler scale, therefore leading to high-complexity receiverprocessing, which is not suited for continuous data monitoringand detection.

Our proposed approach is to use a bank of self-correlators,see Fig 2, with each one matched to a different periodicity.Detection, synchronization, and Doppler scale estimationareaccomplished based on the correlation metrics from the bankof self-correlators.

We now present the proposed receiver processing, basedon the sampled baseband signal. The baseband signal isusually oversampled at a multiple of the system bandwidthts = 1/(λB):

y[n] = y(t)|t=nts, (13)


where the oversampling factorλ is an integer. The receiverprocessing includes the following steps:

1) Each of theL branches calculates a correlation metricwith one candidate value of the window sizeNl. Thewindow sizeNl shall be close toλK0, which is thenumber of samples of one OFDM symbol when noDoppler scaling occurs. For each delayd, the correlationmetric corresponding to the window sizeNl is

M(Nl, d) =∑d+Nl−1

i=d y∗[i] y[i+Nl]√

∑d+Nl−1i=d |y[i]|2 · ∑d+Nl−1

i=d |y[i+Nl]|2, (14)

for l = 1, . . . L.2) A detection is declared if the correlation metric of any

branch exceeds a preset thresholdΓth:

H1 if: maxl

|M(Nl, d)| > Γth (15)

Since the norm of the metric in (14) is between 0 and1, the thresholdΓth takes a value from [0,1].

3) The branch with the largest correlation metric is viewedas having the best match on the repetition length. SinceDoppler scaling changes the periodT0 to T0/(1 + a),the Doppler scale factor can be estimated as

a =λK0 − N

N, whereN = arg max

{Nl}|M(Nl, d)|

(16)The speed estimate follows as

v = ca, (17)

where c is the speed of sound in water. Additionalprocessing can be used to refine the Doppler scaleestimate, as will be shown later on in Section VI.

4) Synchronization is performed on the branch that yieldsthe maximum correlation metric. After the maximumis determined, the start of transmission can be selectedas suggested in [12]; starting from the peak the 80%“shoulders” are found (first sample of this correlatorbranch before and after the peak that is less than 80% ofthe peak) and the middle is chosen as synchronizationpoint. This is beneficial, since due to the CP structure thecorrelation metric has a plateau around the peak [12].

Remark 1 (Doppler scale resolution):Since the windowsizeNl is an integer, the minimum step size on the Dopplerscale is1/(λK0). To improve the Doppler scale resolution,the receiver operates on the oversampled baseband signal.The oversampling factor depends on the needed Doppler scaleresolution and the parameterK0. The maximum value ofλis the ratio of the passband sampling rate to the basebandsignal bandwidth, which is typically less than 50 in underwaterapplications.

Remark 2 (implementation complexity):The three slidingsummations in (14) can be computed recursively. For exam-ple, defineψ(Nl, d) =

∑d+Nl−1i=d y∗[l]y[i + Nl]. Instead of

summing overNl multiplications, one can use

ψ(Nl, d+1) = ψ(Nl, d)+y∗[d+Nl]y[d+2Nl]−y∗[d]y[d+Nl],

(18)which amounts to two complex multiplications and two com-plex additions for each update. Hence, for each delayd,the metricM(Nl, d) in (14) can be computed by no morethan seven complex multiplications, six complex additions,one square root operation, and one division. Note that thecomplexity does not depend on the window size. Such a low-complexity approach was implemented in a DSP board for oneself-correlator branch [17].

Remark 3 (fine-timing):With the estimated Doppler scalea, the receiver can resample the preamble. This way, the“wideband” channel effect of frequency-dependent Dopplershifts can be reduced to the “narrowband” channel effectof frequency-independent Doppler shifts [4]. The fine-timingalgorithms developed for narrowband radio channels, such asthose in [15], [16] can be applied on the resampled preamble.This way, the “first” path can be synchronized [16], instead ofthe “strongest” path in case of the LFM based method. Dueto channel fading, the strongest path can appear at randompositions within the delay spread, which is undesirable forthetask of data block partitioning after the synchronization.Incontrast, the position of the first path is more stable.

So far, we have described the general procedure of the pro-posed detection, synchronization, and Doppler scale estimationalgorithm, while some parameters are left to be specified.Important questions include:

• How to set the detection threshold?• How many parallel branches are needed? What is the

desired Doppler scale resolution?

We next address these questions by analyzing the detectionperformance in Section IV, and the data demodulation perfor-mance under Doppler scale mismatch in Section V.

IV. RECEIVER OPERATING CHARACTERISTICS

The output of each correlator is a random variable dueto the additive noise. The probabilities of false alarmPfa

and detectionPd are the probabilities of the correlator outputexceeding a thresholdΓth under the “no signal” and “signal”hypotheses, respectively. We now analyze the false alarm anddetection probabilities of a single branch, as a function ofthethresholdΓth and the Doppler scalea. This will give us anunderstanding of the necessary Doppler scale spacing in theparallel self-correlator structure for detection purposes. Due toover-sampling, the summations in the metric given in (14) ofthe lth branch can be well approximated with continuous timeintegrals:

M(

T , t)

=

∫ t+T

ty(τ)∗y(τ + T ) dτ

√

∫ t+T

t |y(τ)|2 dτ ·∫ t+T

t

∣

∣

∣y(τ + T )

∣

∣

∣

2

dτ

,

(19)where

T = Nl · ts =T0

1 + a, t = d · ts. (20)


A. Probability of False Alarm

When no signal is present,y(t) = n(t). SinceBT ≈ K0,we can find a set of orthonormal basis functions{fi(t)}K0−1

i=0 ,such that

n(τ) =

K0−1∑

i=0

nt,ifi(τ), τ ∈ [t, t+ T ] (21)

n(τ + T ) =

K0−1∑

i=0

nt+T ,ifi(τ), τ ∈ [t, t+ T ]. (22)

Assume thatn(t) is a Gaussian noise process, thennt,i andnt+T ,i are independent and identically distributed Gaussianrandom variables. Define2

nt = [nt,0, . . . , nt,K0−1]T and

nt+T = [nt+T ,0, . . . , nt+T ,K0−1]T , and their normalized

versions nt = nt/‖nt‖ and nt+T = nt+T /‖nt+T ‖. Thecorrelator output (19) can be simplified to the inner productbetween two unit length vectors as

M(

T , t)

= nHt nt+T . (23)

Finding the probability of false alarm can now be linked tothe Grassmannian line packing problem in [24]. Specifically,nt can be viewed as coordinates of a point on the surfaceof a hypersphere with unit radius, centered around the origin.This point dictates a straight line in a complex spaceCK0 thatpasses through the origin. The two lines generated bynt andnt+T have a distance defined as:

d(nt, nt+T ) := sin(θt) =√

1 − |nHt nt+T |2, (24)

where θt denotes the angle between these two lines. Thedistanced(nt, nt+T ) is known as “chordal distance” [24].Since n(t) is additive white and Gaussian,nt and nt+T

are uniformly distributed on the surface of the hypersphere.Without loss of generality, we can assume thatnt+T isfixed a priori andnt is uniformly distributed, to evaluate thedistribution of the chordal distance. Based on [25, eq. (34)](which was derived based on the geometrical framework in[26]), we infer

Pr{

d2(nt, nt+T ) < z}

= zK0−1, 0 < z < 1. (25)

Hence, the probability of false alarm is

Pfa = Pr{

|M(T , t)| > Γth

}

(26a)

= Pr{

d2(nt, nt+T ) < 1 − Γ2th

}

(26b)

= (1 − Γ2th)K0−1. (26c)

Note thatPfa does not depend on the power of the additivenoise. Once the thresholdΓth is chosen, a constant false alarmrate (CFAR) is achieved independent of the noise level.

2Bold upper case and lower case letters denote matrices and column vectors,respectively;(·)T , (·)∗, and(·)H denote transpose, conjugate, and Hermitiantranspose, respectively.| · | and ‖ · ‖ stand for for the absolute value of acomplex number and the norm of a vector, respectively.

B. Probability of Detection

Assume that the signal is present, we rewrite (10) as

y(t) = xc(t) + n(t), (27)

where the signal is

xc(t) = ej2πafct∑

k∈S

Hks[k]ej2π k

T0(1+a)t

, t ∈ Tcyclic. (28)

Treating the signal as deterministic unknown, we define theautocorrelation function ofxc(t) as

φxx(T,∆T ) =1

T

∫ t+T

t

xc(τ)xc(τ + ∆T )dτ. (29)

The noise is viewed as wide sense stationary, and we defineits autocorrelation function asφnn(τ). Assuming that theintegration is done in a proper window wherexc(t) is welldefined as in (28), we can easily obtain:

E

{

∫ t+T

t

y(τ)∗y(τ + T ) dτ

}

= T φxx(T , T ),

t ∈ Tplateau:=

[

−Tcp − τmax

1 + a,T0(1 + 2a− a)

(1 + a)(1 + a)

]

(30)

E

{

∫ t+T

t

|y(τ)|2dτ}

= E

{

∫ t+T

t

|y(τ + T )|2dτ}

= T [φxx(T , 0) + φnn(0)], t ∈ Tplateau. (31)

In [12], the absolute value of the correlation metric hasbeen approximated as a Gaussian random variable to derivesome approximate results. We would like to follow the sameprinciple here. To this end, we propose to approximate themean of the correlator output as:

E{∣

∣

∣M

(

T , t) ∣

∣

∣

}

≈ T |φxx(T , T )|T φxx(T , 0) + T φnn(0)

=αγ

γ + 1, t ∈ Tplateau, (32)

where γ = φxx(T , 0)/φnn(0) is the signal-to-noise-ratio(SNR) at the receiver, andα is the correlation coefficient

α =|φxx(T , T )|φxx(T , 0)

. (33)

The variance can be approximated as

Var{∣

∣

∣M

(

T , t)

∣

∣

∣

}

≈ 2γ3 + 5γ2 + 3γ + 1

2K0(γ + 1)4, t ∈ Tplateau.

(34)The variance in (34) was derived in the radio channel casewithout Doppler scaling where|α| = 1. We argue that itcan still be used in the case with Doppler scaling, since thevariation of the correlation output is mainly due to additivenoise.


0 0.5 1 1.5 2 2.5 3 3.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

c⋅ε: speed mismatch in m/s

|M(ε

)|

approx.: µapprox.: µ ±2σsim.: µsim: µ ±2σnoise: µnoise: µ +2σ

γ = 10 dB

γ = 0 dB

(a) non-dispersive

0 0.5 1 1.5 2 2.5 3 3.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

c⋅ε: speed mismatch in m/s

|M(ε

)|

approx.: µapprox.: µ ±2σsim.: µsim: µ ±2σnoise: µnoise: µ +2σ

γ = 10 dB

γ = 0 dB

(b) dispersive

Fig. 3. Statistics of the correlator output, including the meanµ and standard deviationσ as a function of the unknown speed for two different levels ofSNR γ; (a) non-dispersive channel with a single path, (b) dispersive channel with an exponential decay profile.

We now specify the correlation coefficientα. Based on (28),we have fort ∈ Tplateau

φxx(T , T ) =1

T

∫ t+T

t

[

x∗c(τ)xc(τ + T )]

dτ (35a)

= e−j2πafcT∑

k∈S

∑

l∈S

Hks[k]H∗l s

∗[l]

× 1

T

∫ t+T

t

ej2π(1+a)

[

kT0

τ− lT0

(τ+T)]

dτ, (35b)

which lead to

|φxx(T , T )| =∣

∣

∣

∑

k∈S

∑

l∈S

Hks[k]H∗l s

∗[l]e−j2π(1+ǫ)l

× ejπ(1+ǫ)(k−l) sinc[(1 + ǫ)(k − l)]∣

∣

∣, (36)

where sinc(x) = sin(πx)/(πx) and(1+ ǫ) = (1+ a)/(1+ a)with a = T0/T − 1 from (20). Therefore,ǫ ≈ a− a. Assumethat ǫ is tiny, we approximate sinc[(1 + ǫ)(k − l)] ≈ δ[k − l].For constant amplitude modulation|s[i]|2 = σ2

s , we have

|φxx(T , T )| ≈ σ2s

∣

∣

∣

∣

∑

k∈S

|Hk|2e−j2πk(a−a)

∣

∣

∣

∣

. (37)

Similar approximation can be done forφxx(T , 0). We thushave

α ≈∣

∣

∑

k∈S |Hk|2e−j2πk(a−a)∣

∣

∑

k∈S |Hk|2. (38)

In summary, withα approximated in (38), the meanµα

approximated in (32), and the varianceσα approximatedin (34), an approximate expression for the probability ofdetection inTplateau is

Pd = Pr{∣

∣

∣M

(

T , t)∣

∣

∣≥ Γth

}

≈ Q

(

Γth − µα

σα

)

, (39)

whereQ(x) = (1/√

2π)∫ ∞

xe−t2/2dt.

C. Numerical Validation

To confirm the theoretical analysis, we use numerical sim-ulation using the following steps:

1) Generate the baseband signal via (10) and sample it asin (13).

2) Compare the statistics of the correlator output at signalstart t = 0 on a branch withNl = λK0.

For the non-dispersive channel with a single path, the meanof the correlator (32) can be simplified to

E{∣

∣

∣M

(

T , 0)∣

∣

∣

}

=αγ

γ + 1= |sinc[K(a− a)]| γ

γ + 1. (40)

The simulation results for the non-dispersive channel areshown in Fig. 3(a); we observe that the loss of correlation dueto the unknown speed is modelled well by the sinc function,while the approximation of the standard deviation is fairlyexact for the low SNR case but only of the right magnitudefor the high SNR case.

For dispersive channels we average over different channelrealizations, where we choose an exponentially decaying chan-nel profile that loses about 20 dB within 10 ms. For eachchannel realization, we evaluate the mean and variance of theGaussian approximation, then average these by approximatingthem via a single Gaussian distribution with matched mo-ments. Since this is basically a Gaussian mixture distribution,the resulting mean is the average mean, while the variance isincreased: it consists of the average variance and the additional“spread” of the means.

The results are in Fig. 3(b); we see an identical behaviorfor K(a− a) ≪ 1 (inside the main lobe of the sinc function).This is because forǫ ≈ a − a = 0, α is fixed as unity (c.f.(38)), while for ǫ 6= 0, α is a random variable dependingon the specific channel realization. Accordingly the additionalvariation inα leads to an increased variance for largerǫ andthe sinc shape is distorted depending on the channel statistics.


10−20

10−15

10−10

10−5

100

0.4

0.5

0.6

0.7

0.8

0.9

1

probability of false alarm

prob

abili

ty o

f det

ectio

n

AWGNapprox.dispersiveapprox. 2

∆ v = 1.5 m/s

∆ v = 2.0 m/s

∆ v = 2.5 m/s

Fig. 4. Receiver operating characteristic (ROC) of the detection scheme forK0 = 512 OFDM carriers,γ = 0 dB and varying speeds; plotted are theGaussian approximation of the probability of detection andthe Monte-Carlosimulated probability of detection (dotted lines) for different channels againstthe exact analytic probability of false alarm.

Still, for a small speed mismatch the behavior can be wellapproximated by the non-dispersive channel.

To assess the effect of Doppler scale mismatch, i.e., un-known speed, on the detection performance, we plot thereceiver operating characteristic (ROC) in Fig. 4. The exactprobability of false alarm (26c) is plotted against the Gaussianapproximation as well as Monte-Carlo simulation results forboth AWGN and dispersive channels. For the non-dispersivecase we see that the simulation results match the Gaussianapproximation reasonably for the chosen speeds and SNR.Comparing to Fig. 3, the detection performance is superbas long as the mean,µα, of the correlator output under thesignal hypothesis and the mean under the noise hypothesisare separated by more than six times the standard deviation,σα.

In case of the dispersive channel, we had seen in Fig. 3that the meanµα of the correlator was always higher forlarge Doppler scales, but at the cost of an increased variance.This results in the ROC’s having less steep slopes, intersectingthe curves for the AWGN channel – accordingly performingbetter for detection probabilities around one half, but worsetowards one (that are the regions of interest). This detrimentaleffect is negligible when the Doppler scale mismatch is lessthan, e.g.,c(a − a) ≈ 1.5 m/s, which are also the regions ofgenerally good performance. Therefore for a limited Dopplerscale mismatch, the detection performance is not changedmuch by the dispersive nature of the channel.

D. Comparison to LFM Based Detection

We now compare the detection performance of our approachwith the detection performance based on an LFM preamblewith matched filter processing at the receiver followed by athreshold test. Although in practice this threshold has to be

estimated online, for this comparison we assume the noiselevel to be known for the LFM based approach.

We use a dispersive channel withL paths, with a uniformpower profile and a total channel length of about 20 ms. Bothapproaches use a bandwidth of 12 kHz and a preamble lengthof about 100 ms. The OFDM based approach hasK0 = 512subcarriers and therefore a detection SNR of approximatelyK0γ, whereγ is the SNR per subcarrier. The LFM uses anupsweep with the same total energy, giving effectively a 3 dBadvantage since the full preamble is used for correlation. Fig. 5shows the ROCs for varying channel conditions and availableoversampling. We note the following:

• The LFM based approach is less favorable when thenumber of paths increases, loosing about 3 dB whenthe number of paths triples; c.f. Fig. 5(a), the ROCfor γ = −3 dB, L = 15 is almost identical with theROC for γ = 0 dB, L = 45. This loss can be linkedto a comparison between selection combining (SC) andmaximum ratio combining (MRC) [27], as the LFMbased approach only utilizes the energy from the strongestpath.

• Compared to the LFM based approach, the multi-carrierapproach degrades faster with SNR, see Fig. 5(a), sincewe use a “noisy” template for correlation. This effectdepends on the per carrier SNRγ. Onceγ is larger than0 dB this effect is negligible.

• The LFM based approach degrades when no oversam-pling is used, see Fig. 5(b), as the width of the ambiguityfunction is small along the time axis.

Note that for communication systems, both methods willwork well in the practical SNR range, as data demodula-tion requires a decent SNR (much higher than the detectionthreshold). Further, the proposed method has a low-complexityimplementation as discussed in Remark 2, which is not avail-able for the LFM based approach. Due to its non-parametricnature and low-cost implementation, the proposed approachis an appealing choice for the preamble design in underwatermulticarrier communication systems.

V. I MPACT OF DOPPLERSCALE ESTIMATION ACCURACY

In this section, we analyze the performance degradation ondata transmission due to Doppler scale mismatch. This willhelp to specify the needed Doppler scale resolution from adata communication perspective.

For data transmission, we use ZP-OFDM. Since block byblock processing is used, let us focus on one ZP-OFDM block.Let T denote the OFDM symbol duration andTg the guardinterval. The total OFDM block duration isT ′ = T + Tg

and the subcarrier spacing is1/T . The kth subcarrier is atfrequency

fk = fc + k/T, k = −K/2, . . . ,K/2 − 1, (41)

wherefc is the carrier frequency andK subcarriers are usedso that the bandwidth isB = K/T . Let s[k] denote theinformation symbol to be transmitted on thekth subcarrier.The non-overlapping sets of active subcarriersSA and nullsubcarriersSN satisfySA∪SN = {−K/2, . . . ,K/2−1}; the


10−80

10−60

10−40

10−20

100

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

PFA

PD

LFM, L = 15LFM, L = 30LFM, L = 45OFDM

γ = 0 dB

γ = −3 dB

(a) multi-path

10−80

10−60

10−40

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

PFA

PD

LFM, λ = 1LFM, λ = 2LFM, λ = 4LFM, λ = 8OFDM

(b) oversampling

Fig. 5. Comparison of ROC between an approach based on an LFM preamble/matched filter processing and our multi-carrier based approach; (a) as thenumber of pathsL increases, the LFM performance decreases,λ = 1; (b) the LFM based approach is also more sensitive to available oversampling,L = 45,γ = 0 dB; OFDM lines for varyingL or λ are have same formatting as LFM case.

null subcarriers are used to facilitate Doppler compensationat the receiver [4]. The transmitted signal in passband is thengiven by

x(t) = Re

{[

∑

k∈SA

s[k]ej2π kT

tg(t)

]

ej2πfct

}

, t ∈ [0, T+Tg],

(42)whereg(t) describes the zero-padding operation, i.e.,g(t) =1, t ∈ [0, T ] and g(t) = 0 otherwise. Due to the adoptedchannel model, the received passband signal is

y(t) = Re

{

∑

p

[

∑

k∈SA

s[k]ej2π kT

(t+at−τp)g(t+ at− τp)

]

×Apej2πfc(t+at−τp)

}

+ n(t), (43)

wheren(t) is the additive noise.A two-step approach to mitigating the channel Doppler

effect was proposed in [4]. The first step is to resampley(t)in the passband with a resampling factorb, which representsour estimate ofa, leading to

z(t) = y( t

1 + b

)

. (44)

Converting to baseband, we obtainz(t)

z(t) = ej2π a−b1+b

fct∑

k∈SA

s[k]ej2π 1+a1+b

kT

t

×[

∑

p

Ape−j2πfkτpg

(

1 + a

1 + bt− τp

)

]

+ n(t), (45)

The second step is to perform fine Doppler shift compen-sation onz(t) to obtainz(t)e−j2πǫt, whereǫ is the estimatedresidual Doppler shift. Performing ZP-OFDM demodulation,

the outputzm on themth subchannel is

zm =1

T

∫ T+Tg

0

z(t)e−j2πǫte−j2π mT

tdt. (46)

Plugging inz(t) and carrying out the integration, we simplifyzm to

zm = C

(

1 + b

1 + a(fm+ ǫ)

)

∑

k∈S

s[k]m,k + vm, (47)

wherevm is the additive noise,C(f) is defined in (8), and

m,k =1 + b

1 + a· sin(πβm,kT )

πβm,kTejπβm,kT , (48)

βm,k = (k −m)1

T+

(a− b)fm − (1 + b)ǫ

1 + a. (49)

Defining the symbol energy asσ2s = E

[

|s[k]|2]

and thenoise variance asσ2

v, we find the effective SNR on themthsubcarrier to be:

γm =|m,m|2σ2

s

σ2v

|C(

1+b1+a (fm + ǫ)

)

|2+

∑

k 6=m

|m,k|2σ2s

. (50)

The first term in the denominator is due to additive noise,while the second term is due to the self-interference arousedby the Doppler scale mismatch. Even when the additive noisediminishes, the effective SNR is bounded by

γm ≤ γm :=|m,m|2

∑

k 6=m |m,k|2(51)

due to self-interference induced by Doppler scale mismatch.We now evaluate the SNR upperbound for two cases:• Case 1: No Doppler shift compensation by settingǫ = 0.• Case 2: Ideal Doppler shift compensation where

ǫopt =a− b

1 + bfc, (52)


−600 −400 −200 0 200 400 6000

10

20

30

40

50

60

70

Subcarrier Index

SN

R (

dB)

∆ v = 0.06 m/s∆ v = 0.2 m/s∆ v = 0.5 m/s

Fig. 6. The SNR upperboundγm for ǫ = ǫopt (thick, full lines) andǫ = 0(thin, dashed lines) as a function of∆v, wherea − b = ∆v/c.

such that

βm,k = (k −m)1

T+a− b

1 + a· mT. (53)

For the first case, the leading term(k−m)/T in βm,k is thefrequency distance between thekth and themth subcarriers,while the second terma−b

1+afm is the extra frequency shift. Forthe second case, the leading term(k −m)/T in βm,k is thefrequency distance between thekth and themth subcarriers,while the second terma−b

1+a · mT is the extra frequency shift.

Since fm is much larger thanmT , we can see that Doppler

shift compensation will improve the performance. Consideran example offc = 27 kHz andB/2 = 6 kHz, we havefm ∈[21, 33] kHz andmaxm

mT = 6 kHz. Hence, the accuracy of

(a− b) can be relaxed at least by four times to reach similarperformance. Doppler shift compensation is one crucial stepin the receiver design, which was the key in the success of thereceivers in [4].

We now numerically evaluate the upperboundγm for ǫ =0 and ǫopt. We setfc = 27 kHz, B = 12 kHz, andK =1024. Fig. 6 shows the bounds for these two cases respectively.Suppose that we want to limit the self noise to be at least20dB below the signal power. In case ofǫ = 0, we need∆v tobe less than0.06 m/s. While in case ofǫopt, the∆v can be aslarge as0.3 m/s.

Fig. 6 provides guidelines on the selection of the Dopplerscale spacing of the parallel correlators. For example, as-suming that the correlator branch closest to the true speedwill yield the maximum metric, then with fine Doppler shiftcompensationǫ = ǫopt we can set the Doppler scale spacingto be 0.4 m/s (where we need∆v to be less than 0.2 m/s) toachieve an SNR upperbound of at least 25 dB. On the otherhand, if an SNR upperbound of 15 dB is sufficient, the Dopplerscale spacing could be as large as 1.0 m/s.

0 0.5 1 1.5 2 2.5 3 3.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

speed in m/s

RM

SE

(v)

coarsemulti−gridinterpolation

Fig. 7. Average velocity estimation error using varying amounts of correlatorsand a simple interpolation between the measured points; thin, dashed lines areγ = 0 dB and thick full lines areγ = 10 dB.

VI. I MPLEMENTATION AND PERFORMANCETESTING

In Section IV, it was shown that forK0 = 512 OFDMcarriers in the preamble, a speed mismatch of up to 1.5 m/swill not degrade the detection performance considerably. Onthe other hand, the SNR analysis for data reception usingK =1024 OFDM subcarriers in Section V indicated that the speedmismatch should not exceed 0.3-0.5 m/s to limit ICI. Thissuggests a multi-grid approach for the implementation:

1) Coarse-grid search for detection. Only a few parallelself-correlators are used to monitor the incoming data.This helps to reduce the receiver complexity.

2) Fine-grid search for data demodulation. After a detec-tion is declared, a set of parallel self-correlators withbetter Doppler scale resolution are used only on thecaptured preamble. Fine-grid search is centered aroundthe Doppler scale estimate from the coarse-grid search.

Instead of multi-grid search, one may also use an inter-polation based approach to improve the estimation accuracybeyond the limit set by the step size. We borrow a techniquefrom [28], which is usually used in spectral peak locationestimation based on a limited amount of DFT samples. Aftercoarse or fine-grid search, let|Xk| denote the amplitude fromthe branch with the largest correlation output, and|Xk−1| and|Xk+1| are the amplitudes from the left and right neighbors.Let ∆a denote the grid spacing. The formula

δ =|Xk+1| − |Xk−1|

4|Xk| − 2|Xk−1| − 2|Xk+1|∆a (54)

can be used to estimate an offsetδ of the Doppler scale devi-ating from the strongest branch towards the second strongestbranch.

A. Simulations for Velocity Estimation

We useK0 = 512 subcarriers and an oversampling factor ofλ = 8 and set the coarse grid spacing as∆a = ∆v/c, where


LFM preamble

LFM postamble

OFDM #1

OFDM #2

OFDM #Nb

OFDM preamble

OFDM postamble

data preamble postamble

Fig. 8. The structure of the data packet used in the Buzzards Bay experiment, Dec. 15, 2006.

0 1 2 3 4 5 60

20

40

60

80

100

120

ms

LFM

Res

ult A

mpl

itude

(a) LFM matched filter

0 50 100 150 200 250 3000

0.2

0.4

0.6

0.8

1

msρ

(b) OFDM self-correlator

Fig. 9. Comparison of the matched filter output for one LFM preamble and the “plateau” output of the proposed synchronization metric.

∆v is 1.46 m/s. Fig. 7 depicts the root mean square error ofthe speed estimatesv = ca, at two SNRs of 0 dB and 10dB. We observe a “saw-tooth” shape for the coarse estimates,and the SNR decrease has little impact on this shape. Thissuggests that the probability of not finding the closest branchis negligible and the dominating error is the quantization tothe coarse grid.

After detection of the coarse-grid search, we use anothersix self-correlators with spacing of∆v = 0.366 m/s to searcharound the estimated Doppler scale from the previous stage.Much improved estimates are obtained, as shown in Fig. 7. Theachieved accuracy exceeds the mismatch specification we setearlier of 0.3-0.5 m/s. We can see more degradation of the saw-tooth shape for low SNR. This is reasonable. As the separationin tentative Doppler scales between correlators diminishes,neighboring correlators will have very similar outputs.

Also, Fig. 7 shows the RMSE for velocity estimation withinterpolation applied after the coarse grid search with∆v =1.46 m/s. We observe that the interpolation approach is veryeffective.

B. Results with Experimental Data

When the transmitter and the receiver are stationary, Dopp-ler scale estimation is not needed, and only one self-correlationbranch is necessary at the receiver. The detection and coarsesynchronization algorithms based on one branch have beenused in the PC- and DSP-based multicarrier modem prototypes[17], [29].

We now work on the data from an experiment performed atBuzzards Bay, Dec. 15, 2006 with a fast-moving transmitter.The same data set was used previously in [4] to demonstrate

the capability of OFDM reception in a dynamic setup. Theused packet structure is shown in Fig. 8, where each packetcontains multiple OFDM blocks. A total of 21 packet trans-missions were conducted when the transmitter was movingtowards the receiver and passing by the receiver at the end.See [4] for the detailed descriptions.

Doppler scale estimation is done in [4] based on the mea-sured time difference between the LFM pre- and postamble.This scheme showed good performance, as after Doppler scalecompensation via resampling the data could be decoded withreasonable BER. The drawback is that the whole packet needsto be buffered before data processing. For example, one packetcontains 32 OFDM blocks withK = 1024 subcarriers wherethe packet duration is about 3.8 s [4]. Buffering 3.8 secondsof data before actual decoding introduces excessive processingdelays.

First we plot a comparison of the timing metrics for amatched filter using the LFM waveform with our proposedscheme in Fig. 9. We observe that the channel energy is fairlyconstrained within the first 2-3 ms – accordingly the plateauobserved at the self-correlator output is almost of lengthTcp,which in this case is 25.6 ms.

We compare the Doppler scale/relative speed estimationaccuracy between the two schemes. In Fig. 10(a) we plot therelative velocity estimates between the sender and receiver;as also an OFDM preamble and postamble were available weinclude both estimates. Even though the postamble would notbe used for decoding purposes as real-time operation is thegoal, it gives intuition about the results, since no ground truthis available. Fig. 10(b) zooms in on the transmissions wherethe Doppler scale changes dynamically. This was caused by


5 10 15 20−5

−4

−3

−2

−1

0

1

2

3

4

5

Transmission

spee

d es

timat

es [m

/s]

LFM pre− & post−amblesOFDM preambleOFDM postamble

(a) full transmission

13 14 15 16 17 18 19 20 21−5

−4

−3

−2

−1

0

1

2

3

4

5

Transmission

spee

d es

timat

es [m

/s]

LFM pre− & post−amblesOFDM preambleOFDM postamble

(b) zoom

Fig. 10. Comparison of velocity estimation techniques; theprevious method calculated the time difference between LFMpre- and post-ambles, while thenew approach is solely based on either preamble or postambleat one particular time.

the transmitter passing by closely at the receiver’s locationaround transmission 19. As the Doppler scale is assumedconstant during each transmission, this will also be the mostchallenging transmissions to decode.

Generally the speed estimates are close, when looking at thedetails (see zoom), it is interesting to consider that the previousmethod uses the duration of a complete transmission while ournew approach is a point estimate. Since the experiment useda large number of OFDM blocks per transmission (32 blocksfor the case of 1024 subcarriers resulting a packet durationof 3.8 s), the time-varying nature of the actual Doppler speedis reflected differently in the two approaches. Inspecting theestimates for transmission 14, 16 or 20, the LFM-based resultis not the average between the OFDM pre- and postamblepoint estimates. We observe that this new proposed methoddiffers from the previous method by no more than1 knot(0.5 m/s) for any transmission.

We next carry out a comparison on the BER performanceusing the data set from [4], where a 16-state rate 2/3 convo-lutional code is used and each OFDM subcarrier is QPSKmodulated. Demodulation and decoding were done twicefor each packet transmission; once using the Doppler scaleestimate obtained from the LFM method and once using theestimate based onthe OFDM preamble only. Fig. 11 showsthat similar uncoded and coded BER results are obtained. Theproposed method, however, does not need to buffer the wholepacket and can start decoding when each new OFDM blockcomes in.

The instances where differences in BERs are observed arewith transmissions 9 and 18. There are several environmentalfactors which could have contributed to this anomaly, in-cluding ship noise and a rapid rate of change in velocity,as discussed in [4]. Perhaps since the original method forvelocity estimation uses the average compression/dilation overthe entire transmission, this new method, which only reliesonone preamble sequence, is more susceptible to rapid changes

2 4 6 8 10 12 14 16 18 20 0

10−4

10−3

10−2

10−1

100

Transmission

BE

R

LFM pre− & post−ambles, uncodedOFDM preamble, uncodedLFM pre− & post−ambles, codedOFDM preamble, coded

Fig. 11. Comparison of the uncoded and coded bit error rates;decoding afterresampling with either the offline speed estimates based on the LFM pre- andpost-ambles or based on the new speed estimate for online processing.

in velocity during a transmission. For scenarios where theDoppler scale changes rapidly, the packet duration can bedecreased, or multiple synchronization blocks are inserted intothe transmission, to allow for more frequent updates of theDoppler scale estimates.

VII. C ONCLUSION

In this paper we proposed a new method for detection,synchronization, and Doppler scale estimation for underwatercommunication based on multicarrier waveforms. We charac-terized the receiver operating characteristic and analyzed theimpact of Doppler scale mismatch on the system performance.Compared to existing LFM-based approaches, the proposedmethod works with a constant detection threshold independent


of the noise level and is suited to handle the presence of densemultipath channels. More importantly, the proposed approachdoes not need to buffer the whole data packet before datademodulation, which is very appealing for the development ofonline realtime receivers for multicarrier underwater acousticcommunication.

REFERENCES

[1] M. Chitre, S. H. Ong, and J. Potter, “Performance of codedOFDM invery shallow water channels and snapping shrimp noise,” inProceedingsof MTS/IEEE OCEANS, vol. 2, 2005, pp. 996–1001.

[2] P. J. Gendron, “Orthogonal frequency division multiplexing with on-off-keying: Noncoherent performance bounds, receiver design and experi-mental results,”U.S. Navy Journal of Underwater Acoustics, vol. 56,no. 2, pp. 267–300, Apr. 2006.

[3] M. Stojanovic, “Low complexity OFDM detector for underwater chan-nels,” in Proc. of MTS/IEEE OCEANS conference, Boston, MA, Sept.18-21, 2006.

[4] B. Li, S. Zhou, M. Stojanovic, L. Freitag, and P. Willett,“Multicarriercommunication over underwater acoustic channels with nonuniformDoppler shifts,”IEEE J. Ocean. Eng., vol. 33, no. 2, Apr. 2008.

[5] W. Li and J. C. Preisig, “Estimation of rapidly time-varying sparsechannels,”IEEE J. Ocean. Eng., vol. 32, no. 4, pp. 927–939, Oct. 2007.

[6] Y. R. Zheng, C. Xiao, T. C. Yang, and W.-B. Yang, “Frequency-domain channel estimation and equalization for single carrier underwateracoustic communications,” inProc. of MTS/IEEE OCEANS conference,Vancouver, BC, Canada, Sept. 29 - Oct. 4, 2007.

[7] D. B. Kilfoyle, J. C. Preisig, and A. B. Baggeroer, “Spatial modulationexperiments in the underwater acoustic channel,”IEEE J. Ocean. Eng.,vol. 30, no. 2, pp. 406–415, Apr. 2005.

[8] S. Roy, T. M. Duman, V. McDonald, and J. G. Proakis, “High rate com-munication for underwater acoustic channels using multiple transmittersand space-time coding: Receiver structures and experimental results,”IEEE J. Ocean. Eng., vol. 32, no. 3, p. 663688, Jul. 2007.

[9] B. Li, S. Zhou, M. Stojanovic, L. Freitag, J. Huang, and P.Willett,“MIMO-OFDM over an underwater acoustic channel,” inProc. ofMTS/IEEE OCEANS conference, Vancouver, BC, Canada, Sept. 29 -Oct. 4, 2007.

[10] D. B. Kilfoyle and A. B. Baggeroer, “The state of the art in underwateracoustic telemetry,”IEEE J. Ocean. Eng., vol. 25, no. 1, pp. 4–27, Jan.2000.

[11] B. S. Sharif, J. Neasham, O. R. Hinton, and A. E. Adams, “Acomputa-tionally efficient Doppler compensation system for underwater acousticcommunications,”IEEE J. Ocean. Eng., vol. 25, no. 1, pp. 52–61, Jan.2000.

[12] T. M. Schmidl and D. C. Cox, “Robust frequency and timingsynchro-nization for OFDM,” IEEE Trans. Commun., vol. 45, no. 12, pp. 1613–1621, Dec. 1997.

[13] H. Minn, V. Bhargava, and K. Letaief, “A robust timing and frequencysynchronization for OFDM systems,”IEEE Trans. Wireless Commun.,vol. 2, no. 4, pp. 822–839, Jul. 2003.

[14] R. D. J. van Nee, G. A. Awater, M. Morikura, H. Takanashi,M. A.Webster, and K. W. Halford, “New high-rate wireless LAN standards,”IEEE Communications Magazine, vol. 37, no. 12, pp. 82–88, Dec. 1999.

[15] E. G. Larsson, G. Liu, J. Li, and G. B. Giannakis, “Joint symbol timingand channel estimation for OFDM based WLANs,”IEEE Commun.Lett., vol. 5, no. 8, pp. 325–327, Aug. 2001.

[16] C. R. Berger, S. Zhou, Z. Tian, and P. Willett, “Performance analysison an MAP fine timing algorithm in UWB multiband OFDM,”IEEETrans. Commun., 2008 (to appear);downloadable at http://www.engr.uconn.edu/˜shengli/BZTW08.pdf.

[17] H. Yan, S. Zhou, Z. Shi, and B. Li, “A DSP implementation ofOFDM acoustic modem,” inProc. of the ACM International Workshopon UnderWater Networks (WUWNet), Montreal, Quebec, Canada, Sep.2007.

[18] J. C. Preisig and G. B. Deane, “Surface wave focusing andacousticcommunications in the surf zone,”The Journal of the Acoustical Societyof America, vol. 116, no. 4, pp. 2067–2080, Oct. 2004.

[19] M. Siderius, M. B. Porter, P. Hursky, and V. McDonald, “ModelingDoppler effects for acoustic communications (A),”The Journal of theAcoustical Society of America, vol. 119, no. 5, p. 3397, May 2006.

[20] L. Weiss, “Wavelets and wideband correlation processing,” IEEE SignalProcessing Magazine, vol. 11, no. 1, pp. 13–32, Jan. 1994.

[21] G. Giunta, “Fast estimators of time delay and Doppler stretch based ondiscrete-time methods,”IEEE Trans. Signal Processing, vol. 46, no. 7,pp. 1785–1797, Jul. 1998.

[22] X. X. Niu, P. C. Ching, and Y. T. Chan, “Wavelet based approachfor joint time delay and Doppler stretch measurements,”IEEE Trans.Aerosp. Electron. Syst., vol. 35, no. 3, pp. 1111 – 1119, Jul. 1997.

[23] L. H. Sibul and L. G. Weiss, “A wideband wavelet based estimatorcorrelator and its properties,”Multidimensional Systems and SignalProcessing, vol. 13, no. 2, pp. 157–186, Apr. 2002.

[24] J. H. Conway, R. H. Hardin, and N. J. A. Sloane, “Packing lines, planes,etc.: Packings in Grassmannian space,”Experimental Math., vol. 5, no. 2,pp. 139–159, 1996.

[25] S. Zhou, Z. Wang, and G. B. Giannakis, “Quantifying the power-losswhen transmit-beamforming relies on finite rate feedback,”IEEE Trans.Wireless Commun., vol. 4, no. 4, pp. 1948–1957, Jul. 2005.

[26] K. Mukkavilli, A. Sabharwal, E. Erkip, and B. Aazhang, “On beam-forming with finite rate feedback in multiple-antenna systems,” IEEETrans. Inform. Theory, vol. 49, no. 10, pp. 2562–2579, Oct. 2003.

[27] M. K. Simon and M.-S. Alouini,Digital Communication over FadingChannels. New York: Wiley, 2000.

[28] E. Jacobsen and P. Kootsookos, “Fast, accurate frequency estimators[DSP Tips & Tricks],” IEEE Signal Processing Magazine, vol. 24, no. 3,pp. 123–125, May 2007.

[29] S. Mason, R. Anstett, N. Anicette, and S. Zhou, “A broadband underwa-ter acoustic modem implementation using coherent OFDM,” inProc. ofNational Conference for Undergraduate Research (NCUR), San Rafael,California, Apr. 2007.

Sean F. Masonwas born in New Haven, Connecti-cut on August 21, 1983. In 2006 he received theB.S. degree in electrical engineering and is currentlyworking toward the M.S. degree, both in electri-cal engineering, at the University of Connecticut(UCONN), Storrs.

For the periods of June through August in 2007and 2008, he was employed at the Naval ResearchLaboratory in Washington D.C. Mr. Mason’s re-search interests lie in the field of signal processingand communications, specifically synchronization

and wideband modulation for underwater acoustic communication. He is anactive member of the honors society Eta Kappa Nu, has served as a reviewerfor the IEEE Transactions on Vehicular Technology and was a session chairfor the 2008 Oceans conference in Quebec City, Canada.

Christian R. Berger (S’05) was born in Heidelberg,Germany, on September 12, 1979. He received theDipl.-Ing. degree in electrical engineering from theUniversitat Karlsruhe (TH), Karlsruhe, Germany in2005. During this degree he also spent a semester atthe National University of Singapore (NUS), Singa-pore, where he took both undergraduate and graduatecourses in electrical engineering. He is currentlyworking towards his Ph.D. degree in electrical engi-neering at the University of Connecticut (UCONN),Storrs.

In the summer of 2006, he was as a visiting scientist at the Sensor Networksand Data Fusion Department of the FGAN Research Institute, Wachtberg,Germany. His research interests lie in the areas of communications and signalprocessing, including distributed estimation in wirelesssensor networks, wire-less positioning and synchronization, underwater acoustic communicationsand networking. Mr. Berger has served as a reviewer for the IEEE Transactionson Signal Processing, Wireless Communications, VehicularTechnology, andAerospacespace and Electronic Systems. In 2008 he was member of thetechnical program committee and session chair for the 11th InternationalConference on Information Fusion in Cologne, Germany.


Shengli Zhou (M’03) received the B.S. degreein 1995 and the M.Sc. degree in 1998, from theUniversity of Science and Technology of China(USTC), Hefei, both in electrical engineering andinformation science. He received his Ph.D. degreein electrical engineering from the University ofMinnesota (UMN), Minneapolis, in 2002. He hasbeen an assistant professor with the Department ofElectrical and Computer Engineering at the Univer-sity of Connecticut (UCONN), Storrs, since 2003.

His general research interests lie in the areasof wireless communications and signal processing. His recent focus is onunderwater acoustic communications and networking. Dr. Zhou has served asan Associate Editor for IEEE Transactions on Wireless Communications fromFeb. 2005 to Jan. 2007. He received the ONR Young Investigator award in2007.

Peter Willett (F’03) received his BASc (Engineer-ing Science) from the University of Toronto in 1982,and his PhD degree from Princeton University in1986. He has been a faculty member at the Univer-sity of Connecticut ever since, and since 1998 hasbeen a Professor. His primary areas of research havebeen statistical signal processing, detection, machinelearning, data fusion and tracking. He has interests inand has published in the areas of change/abnormalitydetection, optical pattern recognition, communica-tions and industrial/security condition monitoring.

He is editor-in-chief for IEEE Transactions on Aerospace and ElectronicSystems, and until recently was associate editor for three active journals:IEEE Transactions on Aerospace and Electronic Systems (forData Fusion andTarget Tracking) and IEEE Transactions on Systems, Man, andCybernetics,parts A and B. He is also associate editor for the IEEE AES Magazine, editorof the AES Magazines periodic Tutorial issues, associate editor for ISIFselectronic Journal of Advances in Information Fusion, and is a member of theeditorial board of IEEEs Signal Processing Magazine. He hasbeen a memberof the IEEE AESS Board of Governors since 2003. He was GeneralCo-Chair (with Stefano Coraluppi) for the 2006 ISIF/IEEE Fusion Conferencein Florence, Italy, Program Co-Chair (with Eugene Santos) for the 2003IEEE Conference on Systems, Man, and Cybernetics in Washington DC, andProgram Co-Chair (with Pramod Varshney) for the 1999 FusionConferencein Sunnyvale.

detection, synchronization, and doppler scale estimation with

Documents