Ultrasonics 42 (2004) 473–478

www.elsevier.com/locate/ultras

An ultrasonic transducer array for velocity measurementin underwater vehicles

P. Boltryk a, M. Hill a,*, A. Keary b, B. Phillips c, H. Robinson b, P. White a

a School of Engineering Sciences, Electromechanical Research Group, University of Southampton, Southampton SO17 1BJ, UKb H Scientific Ltd., Unit 21 Somerset House, Hussar Court, Brambles Business Park, Waterlooville PO7 7SG, UK

c Chelsea Technologies Group, 55 Central Avenue, West Molesey, Surrey KT8 2QZ, UK

Abstract

A correlation velocity log (CVL) is an ultrasonic navigation aid for marine applications, in which velocity is estimated using an

acoustic transmitter and a receiver array. CVLs offer advantages over Doppler velocity logs (DVLs) in many autonomous

underwater vehicle (AUV) applications, since they can achieve high accuracy at low velocities even during hover manoeuvres. DVLs

require narrow beam widths, whilst ideal CVL transmitters have wide beam widths. This gives CVLs the potential to use lower

frequencies thus permitting operation in deeper water, reducing power requirements for the same depth, or allowing the use of

smaller transducers.

Moving patterns in the wavefronts across a 2D receiver array are detected by calculating correlation coefficients between bottom

reflections from consecutive transmitted pulses, across all combinations of receiver pairings. The position of the peak correlation

value, on a surface representing receiver-pairing separations, is proportional to the vessel’s displacement between pulses.

A CVL aimed primarily for AUVs has been developed. Its acoustical and signal processing design has been optimised through

sea trials and computer modelling of the sound field. This computer model is also used to predict how the distribution of the

correlation coefficients varies with distance from the peak position.

Current work seeks to increase the resolution of the peak estimate using surface fitting methods. Numerical simulations suggest

that peak estimation methods significantly improve system precision when compared with simply identifying the position of the

maximum correlation coefficient in the dataset. The peak position may be estimated by fitting a quadratic model to the measured

data using least squares or maximum likelihood estimation. Alternatively, radial basis functions and Gaussian processes successfully

predict the peak position despite variation between individual correlation datasets.

This paper summarises the CVL’s main acoustical features and signal processing techniques and includes results of sea trials

using the device.

� 2004 Elsevier B.V. All rights reserved.

Keywords: Correlation velocity log; Doppler velocity log; Navigation; Peak finding

1. Introduction

Modern maritime vessels are equipped with inte-

grated navigation systems that provide precise and

accurate position and speed estimates. Motion sensors

and positional systems such as the satellite-based globalpositioning system (GPS) are interlinked to provide real-

time navigation information.

In some circumstances, however, GPS is unavailable

for continuous data updates. For submerged vehicles

*Corresponding author. Tel.: +44-23-8059-3075; fax: +44-23-8059-

3053.

E-mail address: m.hill@soton.ac.uk (M. Hill).

0041-624X/$ - see front matter � 2004 Elsevier B.V. All rights reserved.doi:10.1016/j.ultras.2003.12.036

such as autonomous underwater vehicles (AUVs) the

water column acts as a Faraday cage and the onboard

navigation system cannot receive the electromagnetic

GPS signals. In these conditions the AUV must either

surface or allow an antenna to breach the water surface

in order to receive the GPS information. This isimpossible when the water surface is, for example,

blocked by polar ice and may be unacceptable when

stealth is important or the AUV is operating at a deep

level.

AUVs frequently use inertial navigation systems

(INS) to measure accelerations. Velocity is estimated

through integration of the acceleration data and posi-

tion information achieved through a further integration

Fig. 1. A typical two-dimensional transducer array configuration.

474 P. Boltryk et al. / Ultrasonics 42 (2004) 473–478

process. Whilst INS is extremely accurate in the short

term, it is not generally used alone for determining

velocity and position. In the longer term the reliability of

the estimates are affected by bias errors within the

integration calculations. To counteract this problem it isusual to employ a secondary instrument to routinely

zero the bias. Traditionally the well-established Doppler

velocity log (DVL) has been chosen as the secondary

instrument. However, an alternative ultrasonic-based

system called a correlation velocity log (CVL) has sig-

nificant advantages over DVLs for many AUV appli-

cations. This paper compares the characteristics of

DVLs and CVLs for use in typical AUV applications,and summarises the development of a CVL system de-

signed specifically for use in AUVs.

2. Navigation in AUVs

DVLs and CVLs represent two options from a range

of instruments that may be selected for navigating anAUV in conjunction with an INS system. The other

alternatives include mechanical flowmeters, pitot tubes

and electromagnetic speed logs. Since these systems

measure the speed of the vessel through the water their

accuracy may be affected by boundary layer effects and

corrections are necessary to compensate for crosscur-

rents in the seawater. In contrast, both DVLs and CVLs

can track the vehicle speed relative to fixed acousticscatterers on the seabed resulting in higher confidence in

the speed estimate because they reference the stationary

seabed [1].

DVLs have traditionally been the standard choice for

integration into AUVs; CVLs have not affected this

domination of the marketplace because of the high level

of signal processing required for correlating pairs of

signals in CVLs. However, the rise of DSP technologyhas meant that CVLs are now a realistic proposition

with the potential to offer improved velocity estimate

performance over a DVL. Whilst DVLs and CVLs both

operate by insonifying the seabed with ultrasonic signals

there are distinct differences between the operating

principles of the two instruments that significantly af-

fects their performance.

3. CVL operating principle

The specific operating principle of CVLs has been

described in detail elsewhere (see for example, [1–3]). In

its simplest form the CVL consists of an ultrasonic

transmitter and an array of receiving elements, a typical

example of which is shown in Fig. 1. Two short-durationultrasonic signals are transmitted downwards towards

the seabed, separated in time by a known interpulse

interval s. The pulses are reflected by acoustic scatterers

both on the seabed and suspended in the water column.

The waveform detected across the receiving array is

formed by the superposition of all the echo components

from the scatterers. The precise form of this acoustic

wavefront depends on factors such as the reflectivity of

the scatterers and their position in space relative to the

transmitter and receiver array. A CVL moving through

this sound field can therefore distinguish moving pat-terns in the acoustic signature detected across the re-

ceiver array.

There are two specific types of CVL, namely spatial

and temporal CVLs. In a spatial CVL the signals de-

tected by its receivers are compared with those signals

from the second pulse by calculating a correlation

coefficient between all combinations of receivers. Given

a known interpulse interval s the spatial CVL searchesfor the spatial vector D separating the receivers exhib-iting the maximum correlation between consecutive

pulses. In a prototype spatial CVL system described in

this paper, a velocity vector map is constructed by

plotting the spatial vectors separating all combinations

of receiver pairings. The CVL searches for the peak of

the surface generated by superimposing the correlation

coefficient data onto the corresponding velocity vectormap position. Since the interference patterns detected

across the receiving array move at twice the speed of the

CVL through the water, but in the opposite direction,

velocity is estimated using (1).

u ¼ � D2s

ð1Þ

A temporal CVL differs in that it adapts the interpulse

interval s to maximise the calculated correlation coeffi-cient between the acoustic signals detected on a pair of

receivers spaced at a known distance D.Whilst CVL systems detect similarities in the time

histories of received signals across a receiver array, theDVL detects the Doppler shift between transmitted

signals and the resulting backscatter from fixed acoustic

scatterers on the seabed as an indication of vehicle

velocity. To resolve directional information, four inde-

P. Boltryk et al. / Ultrasonics 42 (2004) 473–478 475

pendent narrow beamwidth projectors are often ar-

ranged in a Janus configuration. The Janus configura-

tion arranges the four projectors to point towards the

seafloor in the forwards, aft, port and starboard direc-

tions; differences in the Doppler shift in the four direc-tions can thus be used to estimate velocity. To achieve

the necessary directionality, the ultrasonic transmitters

operate at a relatively high frequency, typically between

100 kHz and 1 MHz.

Fig. 2. Comparing the correlation coefficient distribution as a function

of distance from peak, upper plot using the empirical model to gen-

erate the predicted distribution and lower plot using actual trials data.

4. AUV applications

An important advantage of the CVL system is that

the velocity estimate is independent of sonic velocity

because it can be assumed that during the short inter-

pulse interval separating the pair of transmitted sig-

nals the ocean environment conditions remain constant.

The CVL therefore does not need to make tempera-

ture or salinity profiling measurements to maintain

precision.Whilst DVLs require narrow beams the ideal trans-

mitter for the CVL has a wide beam width, typically

about 60�. Reducing the directionality criterion awayfrom narrow beams allows for smaller transducers than

DVLs for the same operating frequency, or the potential

to use lower frequency operation for the same size of

transducer. This presents two significant advantages for

the application of CVLs to AUV platforms. Firstly,smaller and lighter transducers integrate more easily

with the existing AUV structure. Secondly, operation at

a lower frequency results in lower attenuation of the

acoustic signals. The CVL is either able to operate in

deeper water for the same transmitted power, or is able

to operate with lower power requirements than the DVL

operating in the same conditions. Both CVLs and DVLs

are able to extend their operating range in deeper waterby referencing waterborne scatterers rather than by

being limited to use fixed scatterers on the seabed.

However, this mode of operation is unfavourable be-

cause the velocity estimation is subject to current effects

and therefore the velocity estimate is not necessarily

velocity over ground.

The short-duration pulses transmitted into the water

by CVLs further reduce projector power requirements.Combined with a broadband acoustic signature, the

short pulses are less easily detected and therefore make

the CVL attractive for covert operation.

Another drawback with the DVL is that it requires

there to be a relative velocity between the acoustic tar-

gets and source. It is therefore not well suited to low

speeds or hover situations, which are common charac-

teristic flight profiles for AUVs. In contrast the CVLmaintains accuracy at low speeds and the prototype

system that has been developed is capable of operating

during hover manoeuvres.

5. Development of COVELIA

A spatial CVL system named COVELIA aimed spe-

cifically for AUVs has been developed. Its acoustical

and signal processing design has been optimised throughextensive sea trials and computer modelling of the sound

field.

5.1. Computer modelling and simulation

A detailed model, described in depth in Ref. [5],

investigated the characteristics of the sound field as de-

tected by receivers in the CVL array on a pulse-by-pulsebasis. This detailed model simulates a random distri-

bution of acoustic scatterers on the seabed, each scat-

terer acting as an acoustic source. Using in-depth

models of the acoustic reflectivity characteristics of the

scatterers, the attenuation of the signals through the

seawater and the sensitivity characteristics of the trans-

ducers based on transducer test data, the detailed model

predicts the waveform as detected by each receiver in agiven receiver array configuration. Thorough investiga-

tions into behaviour of the correlation coefficient dis-

tribution across the velocity vector map have developed

an empirical model that is used to predict the distribu-

tion of the correlation coefficient surface as a function of

the distance from the peak position on the velocity

vector map. Parametric tests have investigated the effect

on this correlation coefficient surface distribution causedby changes in receiver array configuration, seabed con-

ditions, vertical velocity, background noise and trans-

ducer properties such as the Q-factor and centre

frequency.

Fig. 2 compares the correlation coefficient surface

distribution as predicted by the empirical model with the

distribution obtained using sea trials of the prototype

476 P. Boltryk et al. / Ultrasonics 42 (2004) 473–478

system. Considering first the trials data in Fig. 2, the

magnitude of the correlation coefficients as calculated

for each receiver are plotted as dots as a function of the

distance from the estimated peak position for each sea

trials dataset. For the modelled data, a peak position isassigned manually at random throughout the velocity

vector map and correlation coefficients assigned to the

measurement points using a binomial distribution

between the limits as defined by the detailed model.

Calculating the absolute distances between the mea-

surement points and the corresponding peak position,

the magnitude of the correlation coefficients can be

plotted as a function of the distance from the peak po-sition. The distribution of the modelled correlation

coefficient data can be illustrated using a large number

of repeat tests. The modelled data here uses parameters

such as depth that are equivalent to the sea trials con-

ditions.

The detailed model has shown that it is preferable to

sample a specific window of the incoming acoustic

waveform rather than calculating correlation coefficientsbased on the entire time history. This correlation time

window (CTW) affects the definition and width of the

correlation coefficient surface’s peak. The earliest re-

turns resulting from specular reflections immediately

below the CVL are found to contain little spatial

information for the CVL. Referring to Fig. 3, progres-

sively delaying the CTW though the entire received

waveform tends to make the correlation peak narrower.However, this also causes the characteristic side lobes of

the correlation surface to become more prominent.

Increasing depth insonifies a wider annulus on the sea-

bed for a given beamwidth and reduces the variance on

the surface.

Fig. 3. The influence of CTW and depth on correlation coefficient

distribution as a function of distance from peak position in element

spacings. Upper three plots vary the CTW for a fixed depth of 40 m.

Lower plots vary depth with CTW set as 20–30% of complete time

history.

A macroscopic model [6] is used to simulate the

performance of COVELIA in the long term using

parameters found using the detailed model. The macro-

scopic model sets up a simulated test run around a user-

defined track, and the AUV is subjected to randomcrosscurrents and waves. Correlation coefficient data is

generated using the empirical model according to the

environmental conditions and the chosen CTW. The

model then predicts velocity and position using peak

finding routines on the simulated correlation data.

5.2. Acoustical development

COVELIA uses a two-dimensional receiving array

that facilitates velocity estimation throughout the for-

wards, backwards, starboard and port directions. It is

additionally possible to infer vertical velocity although

in practice it may be easier to calculate change in depth

using static pressure measurements. The Tonpiltz

transducers are housed in impedance matching oil and

operating at 60 kHz permits operation up to an altitudeof 500 m above the seabed. COVELIA’s 60 kHz oper-

ating frequency is a compromise between a deeper

propagation of the acoustic pulses due to the reduction

in attenuation with decreasing frequency whilst avoiding

the background noise at lower frequencies.

The acoustic power requirements have been mini-

mised to make integration into an AUV favourable. The

transmitted acoustic pulses are very short, typicallyabout 10 acoustic cycles at 60 kHz. Additionally, the

maximum repetition rate is intended to be only about 4

measurement cycles per second in shallow water. Since

the maximum repetition rate is dependent on the height

above the seafloor it is necessary to operate at a lower

repetition frequency in deeper water.

5.3. Onboard signal processing

To ensure that the incoming signals used within the

correlation calculation result from bottom reflections

rather than volume reverberation or background noise,

COVELIA first determines the height of the instrument

above the seabed and uses time of flight calculations to

neglect signals detected before this depth threshold. The

CVL can therefore operate as a depth sounder to verifyother onboard instrumentation. The transmitted power

is then automatically adjusted to take into account the

attenuation of the returns from the seabed caused by

transmission losses and absorption losses by the sea-

floor. After a predetermined post-trigger period, COV-

ELIA samples the incoming received waveforms for a

sample length equal to the required CTW.

The complex form of the correlation coefficient (2)between two complex acoustic signals v1ðtÞ and v2ðtÞ isused by COVELIA [4]. In usual notation, v�2ðtÞ is the

Fig. 4. Comparison of raw COVELIA speed estimate versus Raystar

120 GPS speed prediction.

P. Boltryk et al. / Ultrasonics 42 (2004) 473–478 477

complex conjugate time history of the complex signal

v2ðtÞ.

q ¼Rþ1�1 v1ðtÞv�2ðtÞdt

Rþ1�1 jv1ðtÞj2 dt

Rþ1�1 jv2ðtÞj2 dt

h i1=2 ð2Þ

This allows phase information to be calculated and

the magnitude of the correlation coefficient is found

to be robust to the phase effects caused by vertical

velocity. A signing algorithm is used to improve the

clarity of the correlation surface. The phase of the cor-

relation data at a given velocity vector point is com-pared with the phase at the highest point in the

correlation coefficient dataset. If the difference in phase

exceeds 1=2p then the correlation coefficient magnitudeat that point is multiplied by )1. Referring to Fig. 3,rather than being bounded to lie within the range 0 to 1,

the signed correlation coefficient magnitude varies be-

tween )1 and +1.The spatial vectors separating all combinations of the

receivers are plotted onto a grid called a velocity vector

map. By superimposing the correlation coefficient data

onto this map, a correlation surface is produced whose

peak position represents the optimum estimate for the

separation vector D. The velocity estimate therefore re-quires COVELIA to estimate the peak position of the

surface using the available correlation data. The reso-

lution of the instrument is improved by using peakfinding methods. Various peak estimation methods have

been investigated [7], including those based on highest

point (HP), least squares (LS), and maximum likelihood

estimation (MLE). Current work concentrates on

exploiting the learning behaviour of Gaussian processes

(GP) [8] and radial basis functions (RBFs) [9] for peak

finding.

HP is the most rudimentary peak estimation method,whereby the indices of the highest point in the correla-

tion coefficient dataset are selected as the peak position.

Whilst this method provides no interpolation between

measurement points it has proven to be a reliable fall-

back method when more complicated methods fail. An

axisymmetric quadratic model is fitted to the correlation

coefficient data using both LS and MLE methods. The

correlation surface peak is estimated by finding the peakposition of the fitted model. MLE is implemented in a

non-linear iterative process, which results in weighted

LS, where the weighting function is the expected vari-

ance at the given measurement points. Both GP and

RBF models are trained offline using sample data, where

the inputs to the model are the correlation coefficients

and the output is the predicted peak position. The

trained GP and RBF models make peak predictions inreal-time, and a quality factor may be output from the

model, which is useful for integrating the output data

with other navigational devices using processes such as

Kalman filtering. Simultaneous models are used for

estimating the peak position in the x and y directions.LS is currently the favoured peak estimation method.

It achieves good results throughout the measurement

area, and avoids the iterative process necessary for MLEand the complexity of GPs and RBFs. Preliminary re-

sults suggest that GPs are extremely well suited to peak

finding in this context, although inclusion of this peak

estimation technique could affect the maximum repeti-

tion rate of the measurement cycle due to the additional

computational load over LS.

6. Trials data

COVELIA has undergone a comprehensive test pro-

gramme that has included tests in different conditions. A

reservoir environment has been used extensively, to-

gether with tests offshore. The trials data in this paper

represents a short period during a recent sea trial on a

test boat that is equipped with precision navigationalequipment including a Raystar 120 GPS system.

The peak finding algorithm used during these tests is

LS with HP fallback. Whilst COVELIA estimates

velocity directly, it is more useful here to compare its

performance with respect to the reference GPS system

by considering the speed and the position estimated by

each system. Referring to Fig. 4, the raw COVELIA

speed estimate tracks the GPS speed reliably, althoughthe COVELIA data has a significantly greater variation.

Note that whilst the GPS system uses filtering tech-

niques the COVELIA speed data presented here is

unfiltered. Fig. 5 plots position using GPS data versus

the estimated position using COVELIA over the same

period of about 36 min.

Fig. 5. Comparison of position estimation using Raystar 120 GPS

system and COVELIA.

478 P. Boltryk et al. / Ultrasonics 42 (2004) 473–478

7. Conclusions

DVLs have traditionally been the instrument of

choice for interfacing with INS in the AUV navigation

system. An alternative ultrasonic-based navigational aid

that is aimed specifically for AUVs has been developed

and tested. Rather than detecting Doppler shifts in the

ultrasonic echoes from the seabed, the CVL searches forsimilarities in the signals detected across a receiver

array, and estimates velocity of the vessel based on the

spatial separation vector corresponding to maximum

correlation between receivers, and the time interval

separating a pair of ultrasonic pulses directed at the

seafloor.

COVELIA has been developed using a combination

of computer modelling of the sound field, and a com-prehensive sea trials programme. The computer model-

ling has been influential in the selection of design

parameters such as the optimum correlation time win-

dow. The empirical model of how the correlation surface

distribution varies with distance from peak position has

been used to test a variety of peak finding methods for

improving the precision of the instrument. Whilst LS

combined with capability to revert to highest point has

currently been selected for use in COVELIA, numerical

studies suggest that Gaussian processes offer promisingcapabilities for peak finding on the correlation data.

Sea trials data benchmarked against GPS demon-

strate that the prototype COVELIA is a reliable and

accurate instrument. Current work seeks to fine-tune the

RBF and GP models to make their inclusion into the

device more attractive.

References

[1] B. Denbigh, Ship velocity determination by Doppler and correla-

tion techniques, IEEE Proceedings 31 (1984) 315–326.

[2] B.L. Grose, The application of the correlation sonar to autono-

mous underwater vehicle navigation, in: Proceedings of the 1992

Symposium on Autonomous Underwater Vehicle Technology,

1992, pp. 298–303.

[3] S.K. Hole, B. Woodward, W. Forsythe, Design constraints and

error analysis of the temporal correlation log, IEEE Journal of

Oceanic Engineering 17 (1992) 269–279.

[4] W.S. Burdic, Underwater acoustic system analysis, Prentice-Hall,

Englewood Cliffs, NJ, 1984.

[5] A. Keary, M. Hill, P. White, H. Robinson, Simulation of the

correlation velocity log using a computer based acoustic model,

presented at 11th International Symposium, Unmanned Unteth-

ered Submersible Technology, New Hampshire, 1999.

[6] M. Hill, B. Phillips, H. Robinson, On the development of a

correlation velocity log, presented at International Unmanned

Undersea Vehicle Symposium, Newport, RI, USA, 2000.

[7] P. Boltryk, M. Hill, A. Keary, B. Phillips, H. Robinson, P. White,

Improvement of velocity estimate resolution for a correlation

velocity log using surface fitting methods, presented at Oceans 2002

MTS/IEEE, Biloxi, Mississippi, USA, 2002.

[8] C. Bailer-Jones, H. Bhadeshia, D. MacKay, Gaussian process

modelling of austenite formation in steel, Materials Science And

Technology 15 (1999) 287–294.

[9] M. Orr, Regularization in the selection of radial basis function

centers, Neural Computation 7 (1995) 606–623.