Vol. 36 32 No. No. 232 2019 2014 Vol.
UNDERWATER TECHNOLOGY
ISSN 1756 0543
17
A Personal View... A view from the shallow end
Keith Broughton
19
Impact of underwater bandwidth and SNR on crosscorrelationbased population estimation technique of fish and mammals
Shaik Asif Hossain and Monir Hossen
29
Position estimation for underwater vehicles using unscented Kalman filter with Gaussian process prediction
Wilmer Ariza Ramirez, Zhi Quan Leong, Hung Nguyen and Shantha Gamini Jayasinghe
37
Book Review Project Azorian: The CIA and the Raising of the K129
www.sut.org
0ASUT36(2)OFC.indd 1
7/16/2019 11:49:40 AM
UNDERWATER TECHNOLOGY Editor Dr MDJ Sayer Scottish Association for Marine Science Assistant Editor E Azzopardi SUT Editorial Board Chairman Dr MDJ Sayer Scottish Association for Marine Science Gavin Anthony, GAVINS Ltd Dr MA Atamanand, National Institute of Ocean Technology, India LJ Ayling, Maris International Ltd Commander Nicholas Rodgers FRMetS RN (Rtd) Prof Ying Chen, Zhejiang University Jonathan Colby, Verdant Power Neil Douglas, Viper Innovations Ltd, Prof Fathi H. Ghorbel, Rice University G Griffi ths MBE, Autonomous Analytics Prof C Kuo FRSE, Emeritus Strathclyde University Dr WD Loth, WD Loth & Co Ltd Craig McLean, National Ocean and Atmospheric Administration Dr S Merry, Focus Offshore Ltd Prof Zenon MedinaCetina, Texas A&M University Prof António M. Pascoal, Institute for Systems and Robotics, Lisbon Dr Alexander Phillips, National Oceanography Centre, Southampton Prof WG Price FRS FEng, Emeritus Southampton University Dr R Rayner, Sonardyne International Ltd Roland Rogers CSCi, CMarS, FIMarEST, FSUT Dr Ron Lewis, Memorial University of Newfoundland Prof R Sutton, Emeritus Plymouth University Dr R Venkatesan, National Institute of Ocean Technology, India Prof Zoran Vukić, University of Zagreb Prof P Wadhams, University of Cambridge Cover Image (top): zoonar.com/syrist Cover Image (bottom): Steve Crowther Cover design: Quarto Design/ kate@quartodesign.com
Society for Underwater Technology Underwater Technology is the peerreviewed international journal of the Society for Underwater Technology (SUT). SUT is a multidisciplinary learned society that brings together individuals and organisations with a common interest in underwater technology, ocean science and offshore engineering. It was founded in 1966 and has members in more than 40 countries worldwide, incIuding engineers, scientists, other professionals and students working in these areas. The Society has branches in Aberdeen, London and South of England, and Newcastle in the UK, Perth and Melbourne in Australia, Rio de Janeiro in Brazil, Beijing in China, Kuala Lumpur in Malaysia, Bergen in Norway and Houston in the USA. SUT provides its members with a forum for communication through technical publications, events, branches and specialist interest groups. It also provides registration of specialist subsea engineers, student sponsorship through an Educational Support Fund and careers information. For further information please visit www.sut.org or contact: Society for Underwater Technology 2 John Street, London WC1N 2ES e info@sut.org
Scope and submissions The objectives of Underwater Technology are to inform and acquaint members of the Society for Underwater Technology with current views and new developments in the broad areas of underwater technology, ocean science and offshore engineering. SUT’s interests and the scope of Underwater Technology are interdisciplinary, covering technological aspects and applications of topics including: diving technology and physiology, environmental forces, geology/geotechnics, marine pollution, marine renewable energies, marine resources, oceanography, salvage and decommissioning, subsea systems, underwater robotics, underwater science and underwater vehicle technologies. Underwater Technology carries personal views, technical papers, technical briefings and book reviews. We invite papers and articles covering all aspects of underwater technology. Original papers on new technology, its development and applications, or covering new applications for existing technology, are particularly welcome. All papers submitted for publication are peer reviewed through the Editorial Advisory Board. Submissions should adhere to the journal’s style and layout – please see the Guidelines for Authors available at www.sut.org.uk/journal/default.htm or email elaine.azzopardi@sut.org for further information. While the journal is not ISI rated, SUT will not be charging authors for submissions.
in more than 40 countries worldwide, including over 190 Corporate Members of the Society.
Disclaimer and copyright The Society does not accept responsibility for the technical accuracy of any items published in Underwater Technology or for the opinions expressed in such items. The copyright of any paper published in the journal is retained by the author(s) unless otherwise stated. All authors are supplied with a PDF version of their papers once published. Authors are encouraged to make the PDF version of their papers free to download from their own websites.
Open Access Underwater Technology is available as Open Access. PDF versions of all published papers from Underwater Technology may be accessed via ingentaconnect at www. ingentaconnect.com/content/sut/unwt. All issues from Volume 20 (1995) onwards are available as Open Access. The Society for Underwater Technology also encourages Underwater Technology authors to make their papers available online on their personal and/or institutional websites for Open Access. Through this arrangement, the Society supports the Open Access policy not only in the UK (the Research Councils UK (RCUK) policy) but also the drive towards Open Access in other countries.
Abstracting and indexing Underwater Technology is included in Emerging Sources Citation Index. Additional abstracting and indexing services include American Academy of Underwater Sciences (AAUS) ESlate; Aquatic Sciences and Fisheries Abstracts (Biological Sciences and Living Resources; Ocean Technology, Policy and NonLiving Resources; and Aquatic Pollution and Environmental Policy); Compendex; EBSCO Discovery Service; Fluidex; Geobase; Marine Technology Abstracts; Oceanic Abstracts; Scopus; and WorldCat Discovery Services.
Subscription Subscription to the print version of Underwater Technology is available to nonmembers of the Society at the following rates per volume (single issue rates in brackets). Prices are given in GBP. Accepted methods of payment are cheque or credit card (MasterCard and Visa). Foreign cheques must be in GBP and drawn on a British bank otherwise a currency conversion surcharge is incurred. UK subscription Overseas subscription
£102.00 (£25.50 per issue) £108.00 (£27.00 per issue)
Underwater Technology is also available in electronic format via ingentaconnect as Open Access. To subscribe to the print version of the journal or for more information please email Elaine Azzopardi at elaine.azzopardi@sut.org
Publication and circulation Underwater Technology is published in March, July and November, in four issues per volume. The journal has a circulation of 2,400 copies to SUT members and subscribers
0BSUT36(2)IFC.indd 1
Advertising To book an advert or for more information please contact Elaine Azzopardi at elaine.azzopardi@sut.org
7/22/19 9:02 PM
www.sut.org
A Personal View...
doi:10.3723/ut.36.017 Underwater Technology, Vol. 36, No. 2, pp. 17–18, 2019
A view from the shallow end I had the great fortune of growing up in the Solomon Islands. I had a dugout canoe from the age of seven, we sailed and, as soon as I was allowed, I learned to dive. At 18, I joined the Royal Navy and trained as a mine warfare and clearance diving officer. On leaving the Royal Navy after 10 years, I sailed from the UK to Australia with my older brother and then moved to the Cayman Islands, where I did all of the above again. My point is that I have spent my life in, on or near the sea – I even studied marine biology at university. I now insure equipment designed to go into the sea and am privileged to see some real cuttingedge technology and innovative solutions to difficult problems. However, I would like to share my views on two concepts that are crucial to starting out the right way to prevent operations going wrong, ensuring technology is not defeated by the elements and helping to avoid large, preventable losses.
Seamanship The first and my favourite theme is seamanship. Like seamanlike behaviour, seamanship is not achieved by accident: it demands a combination of skill, experience and hard work. Take, for example, a remotely operated vehicle (ROV). The management of an ROV is an exercise in seamanship. This starts with the choice of the vessel, which must be suitable for the task and big enough to host the ROV, and able to provide the relevant services required to
run the ROV. There is also the positioning of the launch and recovery system, which needs to be far enough away from machinery or infrastructure that could cause damage or disruption. After all, if the system is too close to those great underwater winches, or if thrusters need to be isolated during launching or recovering, problems can be encountered. When it comes to the ROV crew, the accommodation, food and general environment should be as good as possible to allow the crew to operate effectively. Then we have the bridge crew. Modern, dynamicallypositioned vessels have power aplenty, but power alone will not get you out of trouble. The positioning of the vessel relative to tide and wind requires seamanship, and to assist the ROV crews during any ROV operation, the vessel needs to be positioned correctly. We once had an incident when the deck officer did not listen to the ROV crew and allowed the tide to rotate the ROV during recovery, so much so that the ROV spun round more than 10 times and the umbilical parted. Seamanship is an evolving concept too. Part of seamanship is ensuring our vessels are seaworthy. As our vessels become more interconnected and digitalised, we need to be aware of the threat of cyber events – whether directed or as a collateral damage. Deck and ROV crews are busy, with much to concentrate on. They should be able to rely on the ship’s systems, which should be protected as far as possible – both physically and
Keith Broughton was born in Zambia and grew up in the Solomon Islands in the Western Paciﬁc. He served in the Royal Navy for 10 years, where he qualiﬁed as a mine warfare and clearance diving ofﬁcer. He also has a degree in marine biology from Queen Mary College, London. After working for KPMG in the Cayman Islands as a marine consultant, Keith moved into the insurance sector. He worked as an underwriter for Leviathan, a specialist subsea equipment coverholder, for 17 years until Leviathan was bought by Beazley in 2016. He is now the subsea equipment underwriter at Beazley.
from degradation from a cyber event. All too often an accident is put down to crew error, but have we really investigated all the underlying elements when we assume this? When there is an accident, was the final decision really the causative action? Many submariners go by the theory that there are generally seven incidents that lead to an accident. This figure is not scientific, but demonstrates that a series of events are generally behind an accident, rather than one poor decision or mistake.
Being precise The other concept is preciseness, which has elements of seamanship embedded within it. Another way of thinking about this is: ‘having the correct tool for the correct job’. Whilst on my diving course
17
01SUT32636(2)_Color.indd 17
7/22/2019 6:52:40 PM
Keith Broughton. A view from the shallow end
all hell would break loose if we used a shifter (adjustable spanner) on the clearance diving sets. This would have been down to laziness, as we were always provided with the correct tools, and the sets were nonferrous (no one wants a ferrous diving set on when approaching a magnetic mine – trust me). The wrong spanners would shred the nuts. In this business it is certainly true that one size does not fit all. A large workclass vehicle is not the best unit for shallowwater operations. When using an autonomous underwater vehicle (AUV) or glider, length of mission and power available need to reflect the tides or currents and the rate of
change of water density – you can trap a glider below a salinity layer! Technology is a huge help to us, but we do need to point it in the right direction. Commoditisation of services based on the belief that technology will breach the divide between planning and reality, and one size fits all is a very risky pathway to follow.
Closing remarks As insurers, we review risk by looking at the form of the item to be insured, the function it is expected to carry out, and what perils will be encountered. The same ROV and setup may well need a different supervisor and a
different vessel to carry out different jobs. As insurers, we are looking for an engineering solution. A number of years ago, the managing director of a commercial diving operation was summoned to a client’s office and given a presentation by some technical, noncommercial divers espousing the cost saving that could be achieved by using their techniques over standard commercial techniques. The managing director was asked for comment and simply asked, ‘Can you weld, and can you weld with that getup at 100 m?’, and walked out. It was not a seamanlike solution to the diving problem, and certainly not the right tool for the job!
18
01SUT32636(2)_Color.indd 18
7/22/2019 6:52:42 PM
doi:10.3723/ut.36.019 Underwater Technology, Vol. 36, No. 2, pp. 19–27, 2019
Technical Paper
www.sut.org
Impact of underwater bandwidth and SNR on crosscorrelationbased population estimation technique of ﬁsh and mammals Shaik Asif Hossain* and Monir Hossen Department of Electronics and Communication Engineering, Khulna University of Engineering and Technology, Bangladesh Received November 2018; Accepted March 2019
Abstract Efﬁcient population estimation of ﬁsh and mammals is one of the prerequisites of environmental monitoring and research. In order to overcome the limitations of conventional methods for population size estimation of ﬁsh and mammals, a crosscorrelation based passive acoustic technique is proposed as an additional method. However, limited bandwidth of underwater channels poses a challenge during acquisition of ﬁsh signals with this technique. To overcome this problem, proper scaling is a mandatory task. The present study investigates scaling factors for chirp and grunt signals at 0–5 kHz underwater bandwidth. A low signaltonoise ratio (SNR) is shown to be an impediment to obtain an accurate estimation. The present research concludes that estimation with minimum SNR of 26.02 dB performs similarly to estimation without noise. Keywords: population estimation, crosscorrelation, impact of bandwidth, SNR, acoustic sensor
1. Introduction According to a report by World Wide Fund for Nature (WWF), the population size of fish, marine mammals, birds and reptiles fell 49 % between 1970 and 2012; for fish alone, the decline was 50 % (Doyle, 2015). A joint study by the WWF and the Zoological Society of London (ZSL) found that the population size of some commercial fish, e.g. a group including tuna, mackerel and bonito, had fallen by almost 75 % (Doyle, 2015). Lack of early knowledge about the population and diversity of species, as well as haphazard fishing, are the major reasons for this decline. Therefore, proper fish popu* Contact author. Email address: asifruete@gmail.com
lation estimation is a mandatory task. Accurate fish population estimation is also necessary as ecological research and commercial fishery management activities largely depend on it. However, in harsh underwater environments, proper fish population estimation presents challenges. The dynamics of the populations and the harsh conditions of the ocean represent the main difficulties in obtaining accurate data. Fish population estimation techniques are classified as either nonacoustic or acoustic (Hossain et al., 2019). The main nonacoustic techniques are removal method of population estimation; minnow traps; visual sampling technique; environmental DNA technique; and predictionbased macroecological theory. Removal method of population estimation has been applied to estimate small mammal populations; a particular number of kill traps is set for a series of trapping period. Pollock (1991) illustrates the capturerecapture models; band or tag return models; removal or catch per unit models; selective removals or changeinratio models; radio tagging survival models; and nest survival models. Visual census comprising techniques used to estimate reef fish populations, first described in Brock (1954), has been adopted by the longterm monitoring program (LTMP) to assess reef fish population size (Emslie et al., 2018). An attempt to estimate the population size of aquatic species using environmental DNA concentrations in large stream and river ecosystems is presented in Doi et al. (2017). This technique ensures accuracy, but it is a complex and costly method of estimation. In fact, most of the nonacoustic techniques present disadvantages, including their timeconsuming
19
02SUT31136(2)_Color.indd 19
7/22/2019 6:38:18 PM
Hossain and Hossen. Impact of underwater bandwidth and SNR on crosscorrelationbased population estimation technique of fish and mammals
nature; poor accuracy; reliance on largely human interaction; and costly mechanical instruments. Therefore, the present research focused on acoustic techniques for population estimation of fish and mammals. The conventional acoustic techniques are echointegration technique, dualfrequency identification sonar (DIDSON) technique and dualbeam transducer technique. An echointegrator equation relates fish abundance to echo energy integrated over a time gate that corresponds to the depth channel of interest. Parameters include the equivalent beam angle; the expected backscattering crosssection per fish; the equipment sensitivity; and a timevaried gain correction factor (MacLennan, 1990). The DIDSON technique has been used in environmental management (Boswell et al., 2008; Martignac et al., 2015). However, the limitations of this method are automatic dataset recording and low range of the beam detection which decreases its accuracy (Martignac et al., 2015). The aspects of using a narrow widebeam acoustic transducer for estimating fish population are illustrated in Ehrenberg (1974). In this technique, the acoustic pulse is transmitted with a narrow beam and the echo is received on both the narrow and wide beams. However, the acoustic methods present some limitations, such as the requirement of a larger number of fish for estimation, costly electronic instruments and monitoring, and protocol complexity. Consequently, a crosscorrelationbased passive acoustic technique for population estimation of fish and mammals has been proposed (Rana et al., 2014; Hossain and Hossen, 2018; Hossain et al., 2018; Hossain and Hossen, 2019). In this technique, the sounds of vocalising fish and mammals are processed to estimate their population size. This statisticsbased technique resolves several of the disadvantages of conventional techniques, e.g. complexity, reliance on human interaction, timeconsuming nature, sensitivity and high cost. In the crosscorrelationbased passive monitoring techniques proposed in Rana et al. (2014), Hossain and Hossen, (2018) and Hossain et al. (2018), fish signals (chirps) of infinite bandwidth were used for estimation. The impact of limited bandwidth and signaltonoise ratio (SNR) on estimation were not considered. In the present paper, finiteband fish signals are applied to analyse the impact of limited bandwidth of underwater acoustic channels in crosscorrelationbased population estimation technique of fish and mammals. The impact of high and low SNR on estimation is also discussed. The present paper considers two acoustic signals of fish and mammals, chirps and grunts, in order to demonstrate a performance analysis.
2. Crosscorrelationbased population estimation technique of ﬁsh and mammals: a review Vocal fish and mammals produce sounds that commonly comprise lowfrequency pulses that vary in duration, number and repetition rate (Myrberg, 1978). Varying acoustic behaviours of fish and mammals are considered in crosscorrelationbased passive monitoring technique (Hossain et al., 2018), which analyses a particular volume at which vocalising fish and mammals produce acoustic signals as a result of their acoustic behaviours. Transmitted acoustic signals from N fish and mammals are received by two acoustic sensors at different delays, and are calculated at each sensor location to form composite signals. The formulation of crosscorrelation of fish sound is similar to the formulation of crosscorrelation of Gaussian signal examined by Anower (2011) to estimate the population size of marine fish and mammals. The transmitted signals are received by the acoustic sensor and recorded in the associated computer from which crosscorrelation is executed. Transmission and reception of signals are performed for a time frame, called ‘signal length’. Vocalising fish and mammals are considered as the sources of acoustic signals, and N fish and mammals are distributed over the volume of a large sphere with a centre that lies halfway between the acoustic sensors. A distribution of fish and mammals is shown in simulation in Fig 1(a). A constant propagation velocity, called the sound velocity, Sp, is considered in the medium. Two acoustic sensors, H1, H2, and a fish/mammal (i.e. the acoustic source), N1, are measured, as shown in Fig 1(b). The acoustic sensors H1, H2 and the fish/ mammal N1 are located at (x1, y1, z1), (x2, y2, z2) and (x3, y3, z3), respectively. If the distance between the two acoustic sensors is dDBS then: d DBS
(x1 − x 2 )2 + (y1 − y2 )2 + (z1 − z2 )2
(1)
A signal from N1 is S1(t), which is finite in length. Therefore, the signals received by H1 and H2 are: Sr 11(t ) = α11S11(t
11
Sr 12(t ) = α12S12(t
12
)
(2)
)
(3)
where, τ11 = d1/Sp and τ12 = d2/Sp are the corresponding time delays for the signal to reach each acoustic sensor, and α11 and α12 are the attenuations as a result of absorption. Assuming τ1 is the time shift in the crosscorrelation, the crosscorrelation function (CCF) is: C1( )
∫
+∞
−∞ −
Sr 11(t (t )Sr 12(t (t
1
)d τ
(4)
20
02SUT31136(2)_Color.indd 20
7/22/2019 6:38:21 PM
Underwater Technology Vol. 36, No. 2, 2019
1000
++
0 2000
2000
1000
1000 0
0
Coefficient value of CCF
100 2000
Bins, b
80 60 40 20 0 1.0
0.75
0.5
0.25 0 0.25 Distance (m)
0.5
0.75
1.0
(a)
Fig 2: Bins, b, in the crosscorrelation process 2000
predefined sampling rate, SR ; distance between sensors, dDBS; and speed of signal propagation, SP, as described by Anower (2011):
1000
0 2000 2000 1000 0
(b)
Fig 1: (a) A distribution of ﬁsh and mammals, where the two pluses (+) indicate the acoustic sensors; and (b) From a distribution of ﬁsh and mammals in 3D spaces, ﬁsh/ mammal N1 is considered, where H1 and H2 are the acoustic sensors
this takes the form of a delta function, as it is a crosscorrelation of two signals where one signal is the delayed copy of another. To find the CCF for N fish and mammals, the total number of signals received by the acoustic sensors taken from each of the fish/mammals is calculated. As such, the total signals, S r , at sensor t1 H1 is: N
∑
S (t − τ j 1 )
(5)
j1 j
j =1
and the total signals at sensor H2 by S r is: t2 N
Sr 2
∑
2 ×dd DBS SP
SR
−1
1000 0
Sr
b=
j2
S j (t − τ j 2 )
(8)
Direct calculation of these estimation parameters using statistical expression is complex. As a result, the crosscorrelation problem is reframed as a probability problem using the renowned occupancy problem following the binomial probability distribution, where the parameters are population size of fish and mammals, N and 1/b (Anower, 2011). By reframing the estimation parameter, i.e. the ratio of standard deviation to mean (Anower, 2011), R becomes: σ R= = μ
1 1 N × ×(1 − ) b b = b −1 N N b
(9)
where, μ is the mean of the CCF, and σ is the standard deviation of the CCF. R is considered as an estimation parameter since the ratio of two parameters is independent of signal strength (Anower, 2011). A block diagram representation of this process is shown in Fig 3.
(6)
j =1
Fish and mammals
Acoustic sensors
where τ = dDBS /Sp is the time shift in the crosscorrelation. Therefore, the final CCF between the signals at the acoustic sensors is: ( )
∫
+∞
−∞ −
Srt 1(t (t )Srt 2(t (t
)d τ
(7)
In the CCF, bins, b, (as shown in Fig 2) is defined as the place occupied by a delta inside a space comprising a width twice the distance between acoustic sensors; this is determined by the delay difference of the acoustic signals (Anower, 2011). Deltas of equal delay difference are placed in corresponding bins. The number of bins, b, is achieved from the
Acoustic signals
Composite acoustic signals
Fig 3: Block diagram of crosscorrelationbased ﬁsh population estimation technique
21
02SUT31136(2)_Color.indd 21
7/22/2019 6:38:24 PM
Hossain and Hossen. Impact of underwater bandwidth and SNR on crosscorrelationbased population estimation technique of fish and mammals
Whereas crosscorrelationbased passive fishery monitoring techniques consider chirp signal from diverse fish sounds initially, the present research considers two fish signals, chirps and grunts, to present the impact of limited bandwidth and SNR.
3. Fish sound Vocalisations of different species vary with respect to individual parameters such as frequency and amplitude; varying sound types are categorised by name. Common types of sound are chirps, pops, grunts, growls, hoots, whistles and clicks. In the present study, each acoustic signal from individual N fish and mammals in a volume is given its own attenuation owing to spherical spreading (proportional to range squared) and attenuation over the path to the receivers. Croaker fish produce a sound akin to a chirp signal, and some species of whale such as humpback whales (Megaptera novaeangliae; Winn et al., 1979); some dolphin species such as bottlenose dolphins (Caldwell and Caldwell, 1970); and some mammals species such as dugongs (Dugong dugon; Ichikawa et al., 2011) also produce chirplike sounds. From a sound analysis of Plectroglyphidodon lacrymatus and Dascyllus aruanus species of damselfish, it is found that their generated chirps consist of trains of 12–42 short pulses of three to six cycles, with a duration of 0·6 ms –1·27 ms, and the peak frequency varies between 3400 Hz–4100 Hz (Parmentier et al., 2006). Northern searobin (Prionotus carolinus), Southern striped searobin (P. evolans; Fish et al., 1952; Fish, 1954; Moulton, 1956; Fish, 1970) and Black Sea gurnard (Protasov, 1965) can produce clucklike sounds. The cluck generated by Northern searobin (Prionotus carolinus) has a frequency range of 40 Hz–2400 Hz with a duration of 100 ms (Fish et al., 1952; Fish, 1954; Moulton, 1956; Fish, 1970). Japanese gurnard (Chelidonichthys kumu; Bayoumi, 1970); Grey gurnard (Eutrigla gurnardus; Hawkins, 1968); the oyster toadfish (Opsanus tau; Gray and Winn, 1961; Winn, 1964); the Gulf toadfish (Opsanus beta; Thorson and Fine, 2002); and Porichthys notatus nesting males (Brantley and Bass, 1994) produce gruntlike sounds. Haddocks emit grunts lasting less than 75 ms and consisting of 3–4 pulses, whereas the grunts produced by codfish are typically less than 150 ms in duration and consist of around 9 pulses. Grunts are broadband (up to 3 kHz) pulsed sounds lasting approximately 300 ms. Species such as Pollimyrus adspersus, Cichlasoma centrarchu (Schwarz, 2010) produce a growllike sound. The growls are broadband (100 Hz–2 kHz) pulsed sounds, variable in duration, with the typical pulse repetition rate of 25 pulses per second (pps) (Crawford, 1997).
Hoots and pops are sounds heard exclusively in aggressive interactions. Hoots are made by P. isidori (Crawford et al., 1986), P. ballayi (Crawford, 1991) and P. adspersus (Crawford et al., 1997), and are relatively short sounds (30 ms) with frequencies lower than 1 kHz of nearly sinusoidal waveforms. Pops are produced by species of Chromis chromis (Picciulin et al., 2002), Pollimyrus and Gnathonemus petersii (Rigley and Marshall, 1973), and consist of a series of pulse emissions with focal energies up to 2 kHz–3 kHz. Cod (Gadus morhua) can produce clicklike sounds with a peak frequency of 55.95 kHz + 2.22 kHz; peaktopeak duration 50.70 ms + 60.45 ms (Vester et al., 2004). Beluga (Delphinapterus leucas), the bottlenose dolphin (Tursiops truncates; Turl and Penner, 1990) and the sperm whale (Lopatka et al., 2006) produce similar sound signals. A whistle is common among the killer whale (Orcinus orca; Andriolo et al., 2015), some species of dolphin (e.g. tursiops truncates; Constantine et al., 2004) and various other species of mammals. A typical representation of fish sound can be expressed as: ⎡⎪⎧ ⎛( f ⎤ ⎞⎪⎫ f1 )t 2 X (t ) = A cos ⎢⎢⎪⎨2π ⎜⎜⎜ 2 + f t ⎟⎟⎟⎪⎬ + P ⎥⎥ ⎟⎠⎪⎪ 2d ⎢⎣⎪⎪⎩ ⎜⎝ ⎥⎦ ⎭
(10)
where, f1 is the starting frequency in Hz; f2 is the ending frequency in Hz; d is the duration in second; P is the starting phase; and A is the amplitude (Rana et al., 2014; Hossain et al., 2018). Fig 4 shows a simulated form of a chirp signal; Fig 4(a) represents a simple form of chirp with a duration of 1 s, and Fig 4(b) represents a chirp with linear instantaneous frequency deviation, where the chirp is sampled at 1 kHz for 2 s. The instantaneous frequency is 0 at t = 0 and crosses 200 Hz at t = 1 s. The present paper uses the properties of chirps and grunts from several acoustic species to investigate the impact of underwater bandwidth and SNR on crosscorrelationbased population estimation technique of fish and mammals during simulation.
4. Method and parameters In signal processing, crosscorrelation is a measure of similarity between two series as a function of the displacement of one series relative to the other. This is also known as a sliding dot product, or sliding innerproduct. In this method the fish signals are first collected by two acoustic sensors separated by a set distance in the region of interest; the received signals are then calculated at each of the two sensors’ locations; and, finally, these two noise signals are crosscorrelated.
22
02SUT31136(2)_Color.indd 22
7/22/2019 6:38:29 PM
Underwater Technology Vol. 36, No. 2, 2019
1
Amplitude
0.5 0 0.5 1
0
0.1
0.2
0.3
0.4
0.5 0.6 Time (s)
0.7
0.8
0.9
1
(a)
In practical cases, underwater acoustic channels are band limited owing to the frequency dependency of absorption loss. The SNR also fluctuates. Hence, it is often a challenging task to implement the crosscorrelationbased passive fishery monitoring technique. In this section, the impact of limited bandwidth and SNR have been discussed. The value of SNR was gradually increased to investigate the impact of SNR during simulations, using MATLAB software as the programming tool. The parameters used in the simulations are given in Table 1.
Frequency (Hz)
500 400
5. Impact of underwater bandwidth
300 200 100 0
0.2
0.4
0.6
0.8
1 1.2 Time (s)
1.4
1.6
1.8
(b)
Fig 4: Chirp signal from simulation: (a) a simple simulated form; and (b) spectrogram of chirp with linear instantaneous frequency deviation
To investigate the simulation result, similar tasks to the crosscorrelationbased passive monitoring technique from Rana et al. (2014) and Hossain et al. (2018) were performed. Simulations were carried out with acoustic sensors along a line in the centre of a sphere. A uniform random distribution of fish and mammals were considered during the simulations. To ease the simulations, a negligible amount of power difference among the similar acoustic pulses transmitted by each fish/mammal was considered. The present research did not examine the Doppler effect that might occur owing to the movement of fish. The Doppler effect creates a slight variation in the propagation wavelength, producing propagation delay that can affect the placing of balls in bins in the crosscorrelation process and lead to fractionalsample delays (Anower, 2011). However, the effect of fractional samples has no significant effect on estimation.
Frequencydependent absorption loss is responsible for underwater bandwidth limitation (Stojanovic, 2006). With the increase of bandwidth, absorption loss increases, causing the transmission range to decrease. Channel bandwidth restricts the signal bandwidth, which affects the estimation performance; this is owing to limited bandwidth creating sinc function (Alam et al., 2010) instead of delta function of infinite band signal. This produces undesired peaks in the bins, and the CCF and estimation therefore become corrupted. To illustrate the impact of bandwidth, 0 kHz–5 kHz chirp and grunt signals were used in the simulations instead of infinite bandwidth signals; a lowpass filter is preferable under water to avoid unwanted highfrequency attenuation. The ratio of standard deviation to mean CCF for the finite bandwidth was obtained and is denoted by RfiniteBW, and infinite bandwidth case is denoted by RinfiniteBW. It was observed from simulation that RfiniteBW is almost the constant multiple of RinfiniteBW and the mean of those constants is 0.59512. This constant value is treated as scaling factor. Therefore, for chirp signals of fish and mammals, multiplying theoretical infinite bandwidth RinfiniteBW by the mean will give the theoretical approximation of finite bandwidth RfiniteBW as: R finiteBW = 0.59512 ×
b −1 N
(11)
Table 1: Parameters used in MATLAB simulation Parameters
Values
Dimension of the sphere Distance between the equidistant sensors, dDBS Speed of propagation, Sp Sampling rate, SR Absorption coefﬁcient, a Dispersion factor, k Number of bins, b Number of iterations Considered ﬁsh distribution
2000 m 0.5 m 1500 m/s 60 kSa/s 1 dBm–1 0 39 500 Uniform random distribution
23
02SUT31136(2)_Color.indd 23
7/22/2019 6:38:30 PM
Hossain and Hossen. Impact of underwater bandwidth and SNR on crosscorrelationbased population estimation technique of fish and mammals
7
5
Theoretical: infinite bandwidth Simulated: infinite bandwidth Theoretical: finite bandwidth Simulated: finite bandwidth
6 5
4
R of CCF
R of CCF
7
Theoretical: infinite bandwidth Simulated: infinite bandwidth Theoretical: finite bandwidth Simulated: finite bandwidth
6
3
4 3
2 2 1 1 0 0 10
1
10 Number of fish and mammals, N
2
10
Fig 5: R of CCF versus N plot (x log and y normal scale) for ﬁnite (0–5 kHz) and inﬁnite bandwidth case, where b = 39 (dDBS = 0.5 m and SR = 60 kSa/s) for chirp signal
⎛ 0.59512 ⎞⎟ ⎟⎟ ×(b − 1) = ⎜⎜⎜ ⎜⎝ R finiteBW ⎟⎟⎠
0 0 10
1
10 Number of fish and mammals, N
2
10
Fig 6: R of CCF versus N (x log and y normal scale) for ﬁnite (0–5 kHz) and inﬁnite bandwidth case, where b = 39 (dDBS = 0.5 m and SR = 60 kSa/s) for grunt signal
2
N chirrp
(12)
Fig 5 shows results related to chirpgenerating fish and mammals; the x axis is taken in logarithmic scale, and the y axis as normal scale. The black solid line represents the estimation with infinite underwater bandwidth from theory, and the black circles represent estimation with infinite underwater bandwidth from simulation. The grey solid line represents the estimation with 0 kHz–5 kHz underwater bandwidth from theory, and the grey stars represent estimation with 0 kHz–5 kHz underwater bandwidth from simulation. The grunt signal of fish and mammals is calculated as: R finiteBW = 0.55245 ×
b −1 N
⎛ 0.55245 ⎞⎟ ⎟⎟ ×(b − 1) = ⎜⎜⎜ ⎜⎝ R finiteBW ⎟⎟⎠
(13)
6. Impact of SNR It is important to consider the impact of noise on estimation of fish population; signals from fish and mammals must be stronger than the noise in order to be useful. If the signals are received by the acoustic sensors with noise, the CCF will be the result of both noise and signal. Although the effect of noise in the proposed population estimation technique will be similar for many types of noise (assuming additive white Gaussian noise, AWGN), the noise strengths will be different. Therefore, the impact of SNR was analysed for the internal noise of a receiver. In the proposed estimation technique, SNR is used as the ratio of voltage levels of signal and noise. An acoustic signal received by two noisy acoustic sensors can be considered as:
2
N grunt
(14)
The result for gruntgenerating fish and mammals is illustrated in Fig 6; the x axis is taken in logarithmic scale, and the y axis as normal scale. The black solid line represents the estimation with infinite underwater bandwidth from theory and the black circles represent estimation with infinite underwater bandwidth from simulation. The grey solid line represents the estimation with 0 kHz–5 kHz underwater bandwidth from theory, and the grey stars represent estimation with 0 kHz–5 kHz underwater bandwidth from simulation. From analysis of Figs 5 and 6, it is found that multiplying by the scaling factor of 0.59512 for chirp signal, and 0.55245 for grunt signal, provides accurate estimation at finite underwater bandwidth of 0 kHz–5 kHz.
f (t ) = S1(t ) + Sn 1(t )
(15)
f 2(t ) = S2(t ) + Sn 2(t )
(16)
where, S1(t) is the delayed version of the signal transmitted from a fish/mammal to acoustic sensor 1; S2(t) is the delayed version of the signal transmitted from the fish/mammal to acoustic sensor 2; Sn1(t) is the internal noise received in acoustic sensor 1; and Sn2(t) is the internal noise received in acoustic sensor 2. The CCF, C (τ), is: C( )
1 T →∞ 2T lim
T
∫
f1(t ) f 2(t
)dt
−T
T ⎡ 1 T ⎤ 1 ⎢ ⎥ S ( t ) S ( t ) d dt + S ( t ) S ( t − ) dt 1 2 1 n 2 ∫ ∫ ⎢ 2T ⎥ 2T −T ⎢ ⎥ −T = lim ⎢ ⎥ T T T →∞ ⎢ 1 ⎥ 1 ⎢+ ⎥ S ( t ) S ( t − ) dt + S ( t ) S ( t ) dt 2 ∫ n1 n 2 ⎢ 2T ∫ n 1 ⎥ 2 T −T −T ⎣ ⎦ = C S1S2 ( ) + C S1Sn 2 ( ) + C Sn 1S2 ( ) + C Sn 1Sn 2 ( )
(17)
24
02SUT31136(2)_Color.indd 24
7/22/2019 6:38:32 PM
C( ) C S1S2 ( )
(18)
demonstrating how the crosscorrelation works with finite time integration. In order to demonstrate the impact of SNR on the estimation technique, the simulations were investigated by adding white Gaussian noise to the signals in the receivers. SNR was also converted to dB, e.g. SNR = 1 indicates 0 dB, SNR = 10 indicates 20 dB, and so on. In the present research, the internal noise of the acoustic sensors was added to the estimation process. Simulations were conducted for the variation of estimated population size of fish and mammals with respect to the variation of SNR. The simulation parameters were the same as those used in the basic estimation technique. In Fig 7, the x axis is taken in logarithmic scale, and the y axis as normal scale. Figs 7(a) and 7(b) are plotted for chirp and grunt signals of fish and mammals, respectively, to show the impact of SNR on crosscorrelationbased population estimation technique of fish and mammals. In Figs 7(a) and 7(b), the black solid line represents the estimated population size of fish and mammals for 100 fish and mammals without noise; the grey solid and grey dotted lines with circles represent the estimated population size of fish and mammals for 100 similar fish and mammals with different SNR. It can be seen from Fig 7 that the increase of SNR results in an increase in the estimation accuracy. When the SNR is 20, the estimation begins to show similar results to those without noise. The result remains nearly the same with further increase of SNR after the value of SNR = 20. It can therefore be concluded that SNR = 20, or 26.02 dB SNR, is the optimum SNR in crosscorrelationbased passive monitoring technique of fish and mammals.
7. Discussion Owing to limited underwater bandwidth, crosscorrelation of fish signals results in sinc functions (Alam et al., 2010) in lieu of delta functions, which corrupts the CCF and RinfiniteBW. Therefore, scaling with a proper scaling factor is required to obtain
120 100 Estimation without noise
80
Estimation with noise 60 40 20 0 3 10
10
2
10
1
10
0
10
1
10
2
10
3
Signal to noise ratio (SNR) (a) Estimated number of fish and mammals
where, CS1S2(τ) is the CCF of S1(t) with S2(t); CS1Sn2(τ) is the CCF of S1(t) with Sn2(t); CSn1S2(τ) is the CCF of Sn1(t) with S2(t); CSn1Sn2(τ) is the CCF of Sn1(t) with Sn2(t); and τ is the time delay in the crosscorrelation process (Liu et al., 2009). As S1 (t) and Sn2 (t), Sn1 (t) and S2 (t), Sn1 (t) and Sn2 (t) are independent random processes, their CCFs tend to be zero with the integration time extension, and when the integration time is infinity. Thus, equation 17 becomes:
Estimated number of fish and mammals
Underwater Technology Vol. 36, No. 2, 2019
120 100 Estimation without noise
80
Estimation with noise
60 40 20 0 10
3
10
2
10
1
10
0
10
1
10
2
10
3
Signal to noise ratio (SNR) (b)
Fig 7: SNR vs. estimated number of ﬁsh and mammals, (a) chirp signal; and (b) grunt signal
accurate estimation when using the proposed method. The optimum scaling factor for 0 kHz–5 kHz underwater bandwidth was found to be 0.59512 for chirp signals, and 0.55245 for grunt signals; the scaling factor is independent of b. Degrading in estimated populations occurred as a result of low SNR in the estimation area. The impact of SNR on crosscorrelationbased passive fish monitoring technique is illustrated in Table 2. Table 2 represents the simulated results for 100 fish and mammals. In Table 2, N Chirps and N Grunts represent the estimation results of chirp and grunt signals for 100 fish and mammals with respect to different SNRs. Table 2 shows that at different SNRs (i.e. 0.001, 0.01, 0.05, 0.1 and 0.5; 60 dB, 40 dB, 26.02 dB, 20 dB and 20 dB) where signal strengths were less than noise, the estimation performance was not of a sufficient standard. However, with the increase of signal strength, the estimation performance increases. Similarly, estimation errors decrease with the increase of SNR. At 26.02 dB SNR, the estimation shows nearly similar results to estimation without noise. Here, the estimation is 96.711 % of actual quantity, and 99.2 % of estimation without noise. With further increase of SNRs, the estimation remains nearly the same as estimation with 26.02 dB SNR. Similarly, for grunt signal, nearly the same estimation error as noiseless estimation at 26.02 dB SNR is given. In this case, the estimation is 109.14 % of actual quantity, and 102.6 % of estimation without noise. Therefore, 26.02 dB SNR is the optimum SNR in crosscorrelationbased population estimation technique of fish and mammals.
25
02SUT31136(2)_Color.indd 25
7/22/2019 6:38:37 PM
Hossain and Hossen. Impact of underwater bandwidth and SNR on crosscorrelationbased population estimation technique of fish and mammals
Table 2: SNRs and corresponding population size estimation with error calculation for 100 ﬁsh and mammals SNR
NChirps
0.001 or 60 dB 0.01 or 40 dB 0.05 or 26.02 dB 0.1 or 20 dB 0.5 or 6.02 dB 1 or 0 dB 2 or 6.02 dB 10 or 20 dB 20 or 26.02 dB 100 or 40 dB 1000 or 60 dB Without noise
2.567 4.431 6.981 12.860 57.352 82.671 89.553 95.693 96.711 96.742 96.731 97.531
NGrunts
1.278 2.688 8.995 16.347 34.781 73.663 77.119 79.165 109.14 108.39 108.49 106.373
Error metric (Estimated quantity as the percentage of actual quantity) for chirp signal, EChirps % 2.567 % 4.431 % 6.981 % 12.860 % 57.352 % 82.671 % 89.553 % 95.693 % 96.711 % 96.742 % 96.731 % 97.531 %
However, the present work has some limitations, including negligence of multipath interference; assumption that the delays are integer; negligence of spherical spreading; consideration of a negligible amount of power difference among the similar fish sounds; consideration of acoustic sensors to be laid at the middle of the estimation area; and negligence of no delay condition, i.e. when the distance between the fish and two sensors are equal, there will be no delay to meet the crosscorrelation condition.
8. Conclusion Impact of underwater bandwidth and SNR are two key terms in crosscorrelationbased population estimation technique of fish and mammals. Limited underwater acoustic channel bandwidth creates an impediment to utilising infiniteband fish signals. Hence, scaling of signals with proper scaling factor is a mandatory task to estimate an accurate population size using crosscorrelationbased technique. The optimum scaling factors were found to be 0.59512 and 0.55245 for chirp and grunt signals, respectively, at 0 kHz–5 kHz underwater bandwidth. Likewise, accurate population size estimation requires increased SNR. With the decrease of SNR, the estimation performance also decreases, and therefore an appropriate SNR must be maintained. The optimum SNR for both chirp and grunt signals was found to be 26.02 dB. These findings will benefit the estimation of fish populations in practical cases using crosscorrelationbased population estimation technique of fish and mammals.
Acknowledgement The authors thank the Department of Electronics and Communication Engineering of Khulna University of Engineering & Technology for providing computational resources for the present work.
Error metric (Estimated quantity as the percentage of actual quantity) for grunt signal, EGrunts % 1.278 % 2.688 % 8.995 % 16.347 % 34.781 % 73.663 % 77.119 % 79.165 % 109.14 % 108.39 % 108.49 % 106.373 %
References Alam MJ, Frater MR and Huntington EH. (2010). Improving resolution and snr of correlation function with the increase in bandwidth of recorded noise fields during estimation of bottom profile of ocean. In: Proceedings of IEEE Oceans 2010 Sydney, 24–27 May, Sydney, Australia. Andriolo A, Reis SS, Amorim TO, Sucunza F, de Castro FR, Maia YG and Dalla Rosa L. (2015). Killer whale (Orcinus orca) whistles from the western South Atlantic Ocean include high frequency signals. The Journal of the Acoustical Society of America 138:1696–1701. Anower MS. (2011). Estimation using crosscorrelation in a communications network. PhD thesis, The Australian Defence Force Academy, Canberra, Australia. Available at <http://unsworks.unsw.edu.au/fapi/datastream/ unsworks:10838/SOURCE01, last accessed 4 June 2019>. Bayoumi AR. (1970). Underwater sounds of the Japanese gurnard Chelidonichthys kumu. Marine Biology 5: 77–82. Boswell KM, Wilson MP and Cowan Jr, JH. (2008). A semiautomated approach to estimating fish size, abundance, and behavior from dualfrequency identification sonar (DIDSON) data. North American Journal of Fisheries Management 28: 799–807. Brantley RK and Bass AH. (1994). Alternative male spawning tactics and acoustic signals in the plainfin midshipman fish Porichthys notatus Girard (Teleostei, Batrachoididae). Ethology 96: 213–232. Brock VE. (1954). A preliminary report on a method of estimating reef fish populations. The Journal of Wildlife Management 18: 297–308. Caldwell MC and Caldwell DK. (1970). Etiology of the chirp sounds emitted by the Atlantic bottlenose dolphin: a controversial issue. Underwater Naturalist 6: 6–8. Constantine R, Brunton DH and Dennis T. (2004). Dolphinwatching tour boats change bottlenose dolphin (Tursiops truncatus) behaviour. Biological Conservation 117: 299–307. Crawford JD. (1991). Sex recognition by electric cues in a soundproducing mormyrid fish, Pollimyrus isidori (part 1 of 2). Brain, Behavior and Evolution 38: 20–28. Crawford JD. (1997). Hearing and acoustic communication in mormyrid electric fishes. Marine and Freshwater Behaviour and Physiology 29: 65–86. Crawford JD, Hagedorn M and Hopkins CD. (1986). Acoustic communication in an electric fish, Pollimyrus isidori (Mormyridae). Journal of Comparative Physiology A 159: 297–310.
26
02SUT31136(2)_Color.indd 26
7/22/2019 6:38:38 PM
Underwater Technology Vol. 36, No. 2, 2019
Crawford JD, Cook AP and Heberlein AS. (1997). Bioacoustic behavior of African fishes (Mormyridae): potential cues for species and individual recognition in Pollimyrus. The Journal of the Acoustical Society of America 102: 1200–1212. Doi H, Inui R, Akamatsu Y, Kanno K, Yamanaka H, Takahara T and Minamoto T. (2017). Environmental DNA analysis for estimating the abundance and BIOMASS of stream fish. Freshwater Biology 62: 30–39. Doyle A. (2015). Ocean Fish Numbers Cut in Half Since 1970. Scientific American. Available at <https://www.scientificamerican.com/article/oceanfishnumberscutinhalfsince1970, last accessed 4 June 2019>. Ehrenberg J. (1974). Two applications for a dualbeam transducer in hydroacoustic fish assessment systems. In: Proceedings of Ocean ‘74 – IEEE International Conference on Engineering in the Ocean Environment, 21–23 August, Halifax, Nova Scotia. Emslie MJ, Cheal AJ, MacNeil MA, Miller IR and Sweatman HP. (2018). Reef fish communities are spooked by scuba surveys and may take hours to recover. PeerJ 6: e4886. Fish MP. (1954). The character and significance of sound production among fishes of the western North Atlantic, vol. 14. New Haven: Bingham Oceanographic Laboratory, 109pp. Fish MP. (1970). Sounds of western North Atlantic fishes: a reference file of biological underwater sounds. Baltimore, Maryland: The Johns Hopkins University Press, 231pp. Fish MP, Kelsey, Jr AS and Mowbray WH. (1952). Studies on the production of underwater sound by North Atlantic coastal fishes. The Journal of Marine Research 11: 180–193. Gray G and Winn HE. (1961). Reproductive ecology and sound production of the toadfish, Opsanus tau. Ecology 42: 274–282. Hawkins AD. (1968). A study on the sound production by marine fish. PhD thesis, University of Bristol, Bristol, UK. Hossain SA and Hossen M. (2018). Biomass estimation of a popular aquarium fish using an acoustic signal processing technique with three acoustic sensors. In Proceedings of 2018 International Conference on Advancement in Electrical and Electronic Engineering (ICAEEE), 2224 November, Bangladesh, India. Hossain SA, Hossen M and Anower S. (2018). Estimation of damselfish biomass using an acoustic signal processing technique. Journal of Ocean Technology 13: 92–109. Hossain SA and Hossen M. (2019). Population size estimation of chirp and grunt generating fish and mammals using crosscorrelation based technique with three acoustic sensors. Journal of Ocean Engineering and Science. <Available at https://doi.org/10.1016/j.joes.2019.04.003> Hossain SA, Hossen M, Mallik A and Hasan SM. (2019). A technical review on fish population estimation techniques: nonacoustic and acoustic approaches. Akustika 31: 87–03. Ichikawa K, Akamatsu T, Shinke T, Adulyanukosol K and Arai N. (2011). Callback response of dugongs to conspecific chirp playbacks. The Journal of the Acoustical Society of America 129: 3623–3629. Liu J, Liu J, Zhang T and Li J. (2009). Application of crosscorrelation algorithm in radio weak signal detection. In: Proceedings of the Seventh Annual Communication Networks and Services Research Conference, 11–13 May, Moncton, New Brunswick, Canada.
Lopatka M, Adam O, Laplanche C, Motsch JF and Zarzycki J. (2006). Sperm whale click analysis using a recursive timevariant lattice filter. Applied Acoustics 67: 1118–1133. MacLennan DN. (1990). Acoustical measurement of fish abundance. The Journal of the Acoustical Society of America 87: 1–15. Martignac F, Daroux A, Bagliniere JL, Ombredane D and Guillard J. (2015). The use of acoustic cameras in shallow waters: new hydroacoustic tools for monitoring migratory fish population. A review of DIDSON technology. Fish and Fisheries 16: 486–510. Moulton JM. (1956). Influencing the calling of sea robins (Prionotus spp.) with sound. The Biological Bulletin 111: 393–398. Myrberg AA. (1978). Ocean noise and the behavior of marine animals: relationships and implications. In: Fletcher JL and Busnel RG. (eds.) Effects of noise on wildlife. New York: Academic Press, 169–208. Parmentier E, Vandewalle P, Frédérich B and Fine ML. (2006). Sound production in two species of damselfishes (Pomacentridae): Plectroglyphidodon lacrymatus and Dascyllus aruanus. Journal of Fish Biology 69: 491–503. Picciulin M, Costantini M, Hawkins AD and Ferrero EA. (2002). Sound emissions of the Mediterranean damselfish Chromis chromis (Pomacentridae). Bioacoustics 12: 236–238. Pollock KH. (1991). Modeling capture, recapture, and removal statistics for estimation of demographic parameters for fish and wildlife populations: past, present, and future. Journal of the American Statistical Association 86: 225–238. Protasov VR. (1965). Bioakustka Ryb. Moscow: Akademia Nauk. Rana MS, Anower MS, Siraj SN and Haque MI. (2014). A signal processing approach of fish abundance estimation in the sea. In: Proceedings of the 9th International Forum on Strategic Technology (IFOST), 21–23 October, Cox’s Bazar, Bangladesh. Rigley L and Marshall JA. (1973). Sound production by the elephantnose fish, Gnathonemus petersi (Pisces, Mormyridae). Copeia 1: 134–135. Schwarz A. (2010). Sound production and associated behaviour in a cichlid fish, Cichlasoma centrarchus. Ethology 35: 147–156. Stojanovic M. (2006). On the relationship between capacity and distance in an underwater acoustic communication channel. ACM SIGMOBILE Mobile Computing and Communications Review 11: 41–47. Thorson RF and Fine ML. (2002). Crepuscular changes in emission rate and parameters of the boatwhistle advertisement call of the Gulf toadfish, Opsanus beta. Environmental Biology of Fishes 63: 321–331. Turl CW and Penner RH. (1990). Differences in echolocation click patterns of the beluga (Dolphinopterus leucas) and the bottlenose dolphin (Tursiops truncatus). The Journal of the Acoustical Society of America 86: 497–502. Vester HI, Folkow LP and Blix AS. (2004). Click sounds produced by cod (Gadus morhua). The Journal of the Acoustical Society of America 115: 914–919. Winn HE. (1964). The biological significance of fish sounds. Marine Bioacoustics 2: 213–231. Winn HE, Beamish P and Perkins PJ. (1979). Sounds of two entrapped humpback whales (Megaptera novaeangliae) in Newfoundland. Marine Biology 55: 151–155.
27
02SUT31136(2)_Color.indd 27
7/22/2019 6:38:38 PM
SUT Publications The SUT publishes a peerreviewed technical journal Underwater Technology; a quarterly magazine UT2 and e–magazine UT3; a series of conference proceedings Advances in Underwater Technology, Ocean Science and Offshore Engineering and The Operation of Autonomous Underwater Vehicles; and in–house conference proceedings and collected papers from seminars. All SUT books and conference proceedings are available to purchase from the SUT website www.sut.org/publications/booksandconferenceproceedings/ This is a selection of the larger collection of the Society’s books and conference proceedings available to purchase online.
Can a Lobster be an Archaeologist? Quirky Questions and Fascinating Facts about the Underwater World From exploring lost treasure to sea monsters, ocean rubbish and how to build your own ROV, the book is packed with factual and fun illustrated stories.
Offshore Site Investigation and Geotechnics: Integrated Geotechnologies – Present and Future Proceedings of the international conference held in September 2012
Price: £220
Price: £12.99
Order Ref. C42
ISBN 978 0 906940 55 6
Hardback; 2012
Paperback; 2015
674 Pages
ISBN 978 0906940532
152 Pages
Subsea Control and Data Acquisition 2010: Future Technology, Availability and Through Life Changes Guidance Notes for the Planning and Execution of Geophysical and Geotechnical Ground Investigations for Offshore Renewable Energy Developments
Price: £15 ISBN 978 0 906940 54 9 Paperback; 2014
Proceedings of the international conference held in Newcastle, UK, 23 June 2010 Proceedings of the International Conference
Price: £95
2–3 June 2010 Newcastle, UK
Order Ref. C41
SUBSEA CONTROL AND DATA ACQUISITION 2010
ISBN 978 0906940525
Future Technology, Availability and Through Life Challenges
Hardback, 2010 176 Pages
48 Pages
The Operation of Autonomous Underwater Vehicles, Volume One: Recommended Code of Practice for the Operation of Autonomous Marine Vehicles, Second Edition
Price: £75 Order Ref. C40 ISBN 978 0906940518 Paperback, 2009 78 Pages
The Collaborative Autosub Science in Extreme Environments: Workshop on AUV Science in Extreme Environments Proceedings for the international science workshop held at the Scott Polar Research Institute, University of Cambridge, 1113 April 2007
Price: £95 Order Ref. C39 ISBN 978 0906940501 Hardback, 2008 202 Pages
For orders and enquiries, please contact: Cheryl Ince, Society for Underwater Technology, Unit LG7, 1 Quality Court, London WC2A 1HR t +44 (0)20 3440 5535 e cheryl.ince@sut.org
02SUT31136(2)_Color.indd 28
7/22/2019 6:38:38 PM
doi:10.3723/ut.36.029 Underwater Technology, Vol. 36, No. 2, pp. 29–35, 2019
Technical Paper
www.sut.org
Position estimation for underwater vehicles using unscented Kalman ﬁlter with Gaussian process prediction Wilmer Ariza Ramirez*, Zhi Quan Leong, Hung Nguyen and Shantha Gamini Jayasinghe University of Tasmania, Australian Maritime College Received February 2019; Accepted May 2019
Abstract The present paper explores the use of Gaussian processunscented Kalman ﬁlter (GPUKF) algorithm for position estimation of underwater vehicles. GPUKF has a number of advantages over parametric unscented Kalman ﬁlters (UKFs) and Bayesian ﬁlters, such as improved tracking quality and graceful degradation with the increase of model uncertainty. The advantage of Gaussian process (GP) over parametric models is that GP considers noise and uncertainty in model identiﬁcation. These qualities are highly desired for underwater vehicles as the number and quality of sensors available for position estimation are limited. The application of nonparametric models on navigation for underwater vehicles can lead to faster deployment of the platform, reduced costs and better performance than parametric methodologies. In the present study, a REMUS 100 parametric model was employed for the generation of data and internal model in the calculation to compare the performance of an ideal UKF against GPUKF for position estimation. GPUKF demonstrated better performance and robustness in the estimation of vehicle position and state correction compared to the ideal UKF. Keywords: Unscented Kalman ﬁlter, GPUKF, Gaussian process, underwater vehicles Acronymn list: ADCP acoustic doppler current proﬁler AUV autonomous underwater vehicles CFD computational ﬂuid dynamics DVL doppler velocity log DVS doppler velocity sonar EKF extended Kalman ﬁlter EnKF ensemble Kalman ﬁlter FKF fuzzy Kalman ﬁlter GP Gaussian process GPUKF Gaussian processunscented Kalman ﬁlter GPS global positioning system LOS lineofsight NARX nonlinear autoregressive network with exogenous input * Contact author. Email address: wilmer.arizaramirez@utas.edu.au
PID RMSE RPM SI UKF
photoionisation detector rootmeansquare error revolutions per minute system identiﬁcation unscented Kalman ﬁlter
1. Introduction Development of accurate and robust navigation technologies is essential for achieving high performances in underwater environments. As the need for complex missions increases, there is a growing demand for highly accurate localisation of underwater vehicles for navigation and data collection purposes. In comparison to ground and air vehicles, localisation via the global positioning system (GPS) is rarely available under water. Therefore, navigation strategies that are more robust and independent from GPS are needed. Strategies for navigation of underwater vehicles are to integrate the vehicle velocity from an accelerometer, gyroscope or water speed sensor to obtain a new position estimate (Dunlap and Shufeldt, 1969). If a water speed sensor is employed, the position at speeds below 0.3 m/s cannot be established, as the sensor is not capable of measuring it. In the case of inertial navigation systems, the acceleration is integrated twice with respect to time (Kuritsky and Goldstein, 1990); the double integral generates drift in the position result. This generated drift can be corrected using complementary sensors such as doppler velocity sonar (DVS), and acoustic doppler current profiler (ADCP), together with algorithms such as extended Kalman filter (EKF) and unscented Kalman filter (UKF). In 1960, a Kalman filter was introduced as an optimal solution for state estimation from a linear system using a prediction of a physical model 29
03SUT31836(2)_Color.indd 29
7/12/2019 6:57:35 AM
Ramirez et al. Position estimation for underwater vehicles using unscented Kalman filter with Gaussian process prediction
(Kalman, 1960). As most systems are nonlinear, the Kalman filter was modified to be used with a nonlinear system by multiple techniques generating alteration as UKF and EKF. In the case of underwater vehicles, these techniques and their variations are the most popular. Armstrong et al. (2010) show that apart from the system, the EKF learn a calibration bias for the magnetic heading. However, applications that employ EKF have produced more robust and accurate results compared to the UKF (Allotta et al., 2016a; Allotta et al., 2016b; Vio et al., 2016). Despite the positive results from UKF applications for underwater vehicles, the UKF can nevertheless demonstrate poor performance, as its predictive variances can be too small if the sigma points are not placed in the correct locations. Deficient predictive variance will produce observations with heavy weight in the measurement update, which causes the UKF to fit the noise (Turner and Rasmussen, 2010). Other filters proposed for underwater vehicles are the ensemble Kalman filter (EnKF), fuzzy Kalman filter (FKF) and particle filter. The EnKFs represent the distribution of the system state using a random sample, called an ensemble, and replace the covariance matrix with the sample covariance computed from the ensemble (Mandel, 2009). An FKF is a combination of a fuzzy set with the Kalman filter; the fuzzy set is a mathematical technique to define inaccuracies and generate better estimation than other Kalman filters (Loebis et al., 2003). Ngatini et al. (2017) compare the EnKF and FKF for underwater vehicles and show that the FKF exhibits better results than the EnKF. The particle filter uses a different approach to the EKF by implementing Bayesian filtering. It makes an approximation of the posterior by using a finite number of particles that represent points in the solution space. Each particle is assigned a weight, and the weighted sample points correspond to the solution of the posterior of the particle state. These particles are propagated according to the dynamics of the posterior, and the weight is modified based on support from the likelihood. The advantage of particle filters is that they do not require a state error Gaussian approximation. Despite research to increase the particle filter speeds (Telles da Silva Vale et al., 2015), the computational cost of running such algorithms is too high for an underwater vehicle’s internal computers. The principal disadvantage of these filters is that their performance depends on the accuracy of the model. The calculation of coefficients from mathematical models for underwater vehicles is a complex task that requires a series of experiments (Bishop and Parkinson, 1970), or computational fluid dynamics (CFD) simulations (Zhang et al., 2010).
The quantity of data required is extended if such calculation or simulations are done within commercial vehicles, which are modular and reconfigurable. The calculation of coefficients is complex, as some coefficients are highly sensitive, and an incorrect calculation can reduce the fidelity of the predicted vehicle motion (Sen, 2000). A solution to avoid the calculation of coefficients for a mathematical model is to use nonparametric methods. Ko et al. (2007) introduced the Gaussian processunscented Kalman filter (GPUKF), a modification of the standard UKF with the replacement of parametric models of state, and measurement by nonparametric models obtained from a series of experimental tests. The nonparametric models give a future state prediction and measurement, and the covariance matrices Q of process and R of measurement noises. The nonparametric model learns over a series of real experiments; it therefore includes more nonlinearities than other common methods to characterise the true signal over a series of noisy samples, via the integrated smoothing function of the Gaussian process (GP). The present paper outlines research into the capability of GPUKF to predict and correct the measured states of an underwater vehicle. The required sample frequency and minimum training data proportion for the GPUKF is also presented. A Simulink model of a REMUS 100 was used to produce the training data required for the nonparametric system identification and test the navigation algorithm. An ideal UKF and rootmeansquare error (RMSE) were employed as comparison measures.
2. Underwater vehicle mathematical model Fossen (1994) showed that the nonlinear dynamic equations of motion of an underwater vehicle can be expressed in vector notation. This is defined by a state vector comprising the vector v of velocities on the body frame of the form [u, v, w, p, q, r]T, and the vector η of position in the earthfixed frame (Fig 1) of the form [x, y, z, φ, θ, ψ]T, such that: Mv + C ( v ) v + D ( v ) v + g ( η) = τ
(1)
Fig 1: Underwater vehicle reference frames, vehicle frame is at centre of buoyancy
30
03SUT31836(2)_Color.indd 30
7/12/2019 6:57:37 AM
Underwater Technology Vol. 36, No. 2, 2019
with the kinematic equation: η = J ( η) v
(2)
where η is vector of position and orientation of the vehicle in the earthfixed frame; v is vector of linear and angular vehicle velocities in body fixed frame; v̇ is vector of linear and angular vehicle accelerations in body fixed frame; M is matrix of inertial terms; C(v) is matrix of Coriolis and centripetal terms; D(v) is matrix consisting of damping or drag terms; g(η) is vector of restoring forces and moments owing to gravity and buoyancy; τ is vector of control and external forces; J(η) is rotation matrix that converts velocities in a bodyfixed frame v to an Earthfixed frame velocity η. Equation 1 can be expanded into a more general equation of motion (Gertler and Hagen, 1967; Prestero, 2001). This expansion is a system of six equations with 73 hydrodynamic coefficients.
3. UKF Table 1 shows the basic structure of the UKF that estimates the states of a dynamic system based on a series of observations and internal model. If xk is the state of the system, uk is the control input and zk is the observation at time k, it can be assumed that the dynamic system evolves according to a state transition function, f( ), and an observation function, h( ), such that: xk
f (xk
u k ) + εk
y
h ( x k ) + δk
(3)
where εk is additive with zeromean Gaussian noise with covariance Q k, and δk is the additive observation noise with covariance Rk. The functions f( ) and h( ) are nonlinear, even when the estimate of the state Xk–1 is Gaussian; the estimate after passing the states through the transition function f( ) is no longer Gaussian. To estimate posteriors over the state space model, the UKF requires a stochastic approximation known as the unscented transform Table 1: UKF algorithm
UKF ( xˆ k

,P
k −
, uk , z k , f ( )), (⋅))
〉Prediction 1 : xˆ k , P k −1 ← UT ( xˆ k 2 : Pkk–1 ← Pkk–1 + Q 〉Correction 3 : zˆ k ,S ,C UT xx̂ˆ −1 k
,P

, uk− , (⋅))
( kk− , Pkk− , h( )) ( z−k1 z k k )
Sk + Ck S 4 : Sk 5 : xˆ k ← xˆ k −1 + Ck Sk ( z k 6 : Pk k end
k − k
−1 k
Pkk −1 + Ck S C
T k
zˆ kk −1 )
(Uhlmann, 1995). The unscented transform works by calculating a set of sigma points that are transformed through the nonlinear functions and their respective Gaussian distribution.
4. Regression with GPs A GP is a nonparametric tool capable of learning regression functions from discrete training data. Benefits of GPs include model flexibility, and the abilities to provide uncertainty estimates and learn noise and smoothness parameters from training data (Rasmussen, 2004). A GP represents the posterior distributions over functions based on training data (Ebden, 2008). It assumes that the data is derived from a noisy process of the form: y
f ( i )+ ε
(4)
where ε is a zeromean additive Gaussian noise with variance σn2. A test input x*, conditioned in a set of data 〈x, y〉 will produce a Gaussian distribution with mean: K( X, X, )[ ( X, ) + σn2 I]1 y
y X y ,x ,x*
(5)
and variance: cov(y* )
k ( , x * ) − k (x * , )[K( , X) + σn2 I]] 1 ( , x * )
(6)
where k(x*, x*) is the evaluation of the kernel with respect to the test point x*; k(x*, X) is a vector defined by kernel values between x* and the training inputs; K(X, X) is the square kernel matrix of the training input values. The prediction uncertainty captured by the variance depends on the process noise and the correlation between the test input and training data. The kernel function selection is governed by application; the most widely used is the squared exponential, or Gaussian kernel, which is considered a universal kernel: kSE(x , x ′)
2
⎛ (x x ′)2 ⎞⎟ ⎟ exp ⎜⎜− ⎜⎝ 2 2 ⎟⎟⎠
(7)
where σ2 controls the average distance of the function away from its mean. The length scale, l, determines the twist length in the function. There are two principal methods for learning the hyperparameters Θ, which are Bayesian model interference and marginal likelihood. Bayesian inference assumes that prior data of the unknown function to be mapped is known. A posterior distribution over the function is refined by incorporation of observations. The marginal likelihood method is based on the aspect that some hyperparameters are
31
03SUT31836(2)_Color.indd 31
7/12/2019 6:57:37 AM
Ramirez et al. Position estimation for underwater vehicles using unscented Kalman filter with Gaussian process prediction
more noticeable in their effect over the posterior distribution. Over this base the posterior distribution of hyperparameters can be described with a unimodal, narrow Gaussian distribution (Rasmussen, 2004). The learning of GPs’ hyperparameters Θ is normally achieved by maximisation of the marginal likelihood. The marginal likelihood can be expressed as: p( 
Θ)
1 (
N
)2 K
1 2
e
1 − y T K −1 y 2
(8)
)−
1 2
T
GP −UKF UKF ( x
−1 1
−
N log ( 2
)
u
z GP
f ⋅ GP − h ⋅ )
(
4 : S k ← Sk + Rk −1 5 : xˆ k ← xˆ k −1 + C k S k ( z k −1 k
Pkk −1 + C k S C
(⋅))
)
zˆ kk −1 )
T k
end
the current state. Consequently, if the calculation is outside the identified region, the GPUKF produces higher uncertainty estimates, reflecting the higher uncertainty in the underlying process model.
(9)
To find a solution for the maximisation of loglikelihood multiples, methods of optimisation can be applied, e.g. particle swarm optimisation, genetic algorithms or gradient descent. For deterministic optimisation methods, the computation of likelihood of partial derivatives with respect to each hyperparameter is needed. According to Williams and Rasmussen (2006), loglikelihood derivatives for each hyperparameter can be calculated by: ⎛ ∂L ( ) 1 ∂K ⎞⎟ = − trace ⎜⎜K −1 ⎟ ∂Θi ∂Θi ⎟⎠ ⎝⎜ 1 ∂K −1 + y T K −1 K 2 ∂Θi
P
〉Prediction 1 : xˆ k ,P k −1,Q k (xx̂ˆ k−  ,P k ,,uuk− , 2 : Pkk–1 ← Pkk–1 + Qk 〉Correction 3 : zˆ k ,S ,C ,R ← UT xˆ k , Pk k − ,GP − h( )
6 : Pk k
where N is the number of input learning data points and y is a vector of learning output data of the form [y1; y2;…yN]. To reduce the calculation complexity, it is preferred to use the logarithmical marginal likelihood obtained by the application of logarithmic properties to equation 8: 1 L (Θ) = − log ( 2
Table 2: GPUKF algorithm
(10)
5. GPUKF The objective of the GPUKF is to replace the internal parametric model f used for state calculation and observation model h with a nonparametric model generated by GPs, and to use the respective variance for the calculation of Qk and Rk. The process noise covariance is obtained from the predictive GP uncertainty at the previous mean sigma point and used for the calculation of the sigma points. The sigma points are passed through the GP observation model, and the observation error covariance is obtained from the observation GP. Table 2 shows the basic structure of the GPUKF algorithm. The incorporation of GP regression allows GPUKFs to learn their models and noise processes from training data. The noise models of the filter automatically adapt to the system states depending on the density of training data around
6. Test setup and results A simulation model of a REMUS 100 (Fig 2), based on the work of Prestero (2001) and Hall and Anstee (2011), was developed in the MATLAB/Simulink software environment and employed to produce data for test and training of the GP. A block diagram of the REMUS 100 model is shown in Fig 2. A pathfollowing controller (Xiang et al., 2017) composed of a lineofsight (LOS) law that pursues a point P(t) and three robust photoionisation detector (PID) controllers, produces the signals for revolutions per minute (RPM), elevator force and rudder force required to control the vehicle. The controllers employ the corrupted measurement to calculate the required forces. A sample frequency of 5 Hz was used to capture data, and a subsample of 40 % of the data was taken randomly for the training. The training data has more points at the start of the trajectory, and the quantity of points reduces over time. Fig 3 shows the selected data for training compared to the simulation data. The virtual sensors employed were a 3axis gyroscope, 3axis accelerometer, compass and doppler velocity log (DVL) unit; the measurement results produce the vector [u,v,w,p,q,r,Z,θ,ψ,ϕ]. Each sensor was simulated by a model comprising an additive noise source and digitalisation of the measurement through a 12bit ADCP. A helix movement was employed to capture 800. The AUV was accelerated from an initial velocity of 0.5 m/s to 1.4 m/s. Fig 4 shows the recorded command signals for 800 s. The noise in the depth sensor required a hard response by the integral parts of the PID, causing the vehicle to converge onto the desired path. A total of 20 simulations were carried out to
32
03SUT31836(2)_Color.indd 32
7/12/2019 6:57:45 AM
Underwater Technology Vol. 36, No. 2, 2019
Fig 2: REMUS 100 simulation model; Xv,Yv,Zv are the vehicle position
u 1.4 Measurement Real state UKF GPUKF
1.2 1 0.8 0.6 0.4 0.2 0
Fig 3: Training and evaluation data
10
20
30
40
50 Time(s)
60
70
80
90
100
Fig 5: Surge speed comparison between corrupted measure data, real position, UKF and GPUKF X 80
Real state UKF GPUKF
60 40 20 0 20 40 60
Fig 4: Helix test input signals to control surfaces
capture data. The first set of data was employed for the creation of the nonparametric GP model. The state vector was defined as the combination of vehicle speeds and vehicle position η = [u,v,w,p,q,r,X,Y,Z,θ,ψ,ϕ]. A UKF was also implemented as an evaluation measure; the filter uses the original REMUS 100 model from which the data was captured to allow comparison of the GPUKF to an ideal UKF when all parameters from the vehicle are known. The GPUKF and UKF were required to estimate the x and y states from the vehicle. The algorithms of Deisenroth et al. (2009) were employed with minor modification to the GP to allow a nonlinear autoregressive network with exogenous input (NARX) structure to be utilised for system identification (SI). The modification included the assembly of the input vector for learning as Xd = [η,ui], where η is the state vector and ui is the command signal. The output vector is formed from the delay vector Y d = Xd(k − 1,k − 2).
80 0
100
200
300
400 Time(s)
500
600
700
800
Fig 6: Predicted position state x comparison between UKF and GPUKF
Fig 5 shows the comparison between the measured data of the real vehicle state, UKF and GPUKF. The GPUKF is equally capable of correcting the measurement state of the vehicle, as the positioning of the sigma points is estimated from the GP’s dynamic model. The predicted position states (Figs 6 and 7) from the GPUKF x and y have a higher similarity to the vehicle’s real position; although the GP employs data corrupted by noise, it has learned to predict over this data. The comparison of the vehicle position estimations and real position is shown in Fig 8. The UKF shows a drift in the calculation of x and y over time. In comparison to the UKF, the GPUKF demonstrates better performance in the prediction of the vehicle position for both the horizon of the training data from the GP, and the decay outside the
33
03SUT31836(2)_Color.indd 33
7/12/2019 6:57:51 AM
Ramirez et al. Position estimation for underwater vehicles using unscented Kalman filter with Gaussian process prediction
Y 80
Real state UKF GPUKF Start
Real state UKF GPUKF
60 40 20 0 20 40 60 80 0
100
200
300
400 Time(s)
500
600
700
800
Fig 7: Predicted position state y comparison between UKF and GPUKF
Fig 8: Comparison of position estimation between real state, UKF and GPUKF
Table 3: Mean RMSE and standard deviation results from UKF and GPUKF per state for 20 runs u(m/s)
v(m/s)
w(m/s)
p(rad/s)
q(rad/s)
r(rad/s)
UKF σUKF GPUKF σGPUKF
0.0014 0.0003 0.0158 0.0387 X(m)
0.0015 0.0021 0.0369 0.0069 Y(m)
0.0005 0.0008 0.0154 0.0033 Z(m)
0.0023 0.0078 0.0569 0.0123 φ(rad)
0.0013 0.0035 0.0601 0.0130 θ(rad)
0.0031 0.0121 0,1813 0,0347 ψ(rad)
UKF σUKF GPUKF σGPUKF
2.4237 14.6908 2.0541 11.1893
1.2695 8.5834 1.3894 9.9445
0.0017 0.0019 0.0281 0.1889
0.0006 0.0020 0.0235 0.0047
0.0012 0.0016 0.0297 0.0052
0.0013 0.0988 0.1095 0.3848
Table 4: Mean RMSE results from UKF and GPUKF for correction and prediction RMSE
Value
RMSEUKF correction RMSEGPUKF correction RMSEUKF prediction RMSEGPUKF prediction
0.0028 0.086 2.91 2.86
training horizon. Although the error increases outside the training horizon, this is corrected by the filter in the return to the training horizon. Tables 3 and 4 summarise the measurement of the mean RMSE between the real vehicle states and the correction from the UKF and the GPUKF for the 20 simulations. The results confirm that the GPUKF can perform as reliably as an ideal UKF, and in some cases it can overperform the ideal UKF. Table 4 shows the average measurement for correction and prediction. In the case of prediction, the GPUKF can forecast vehicle position with higher precision than an ideal UKF, as the inclusion of the noise function and smooth prediction of the GP supports the ability of the UKF to predict position.
7. Conclusion The present research demonstrates that the GPUKF is a promising approach for state estimation in applications where accurate parametric model is
not available. The GPUKF shows similar performance compared to an ideal UKF in the prediction and correction of the vehicle states for the helix movement test case. The average RMSE (as shown in Table 4) for prediction and correction shows that nonparametric models can be employed as prediction models inside the Kalman filter as the UKF for autonomous underwater vehicles (AUVs). The GPUKF demonstrates better performance in the prediction of states than the UKF. The smoothing kernel of the GP facilitated a smooth transition between the prediction points and better placement of sigma points. The tuning complexity in the implementation of a nonlinear Kalman filter is reduced dramatically, as the user is not required to produce the covariance matrices for the process and measure noise model. The GPUKF can be converted to an important tool for underwater vehicles, especially in cases where high nonlinearities are expected, such as in operations near surface or near another object, or during specialised missions. Another advantage of GPUKF is that it can be used even when no GPS signal is available, which is essential for correction in traditional Kalman filters. An underwater vehicle can switch between filters as the availability of data is reduced. The principal disadvantage of GP models is their computational cost during training. Nonetheless, research has shown that this cost can be reduced by using sparse GP models, thus allowing the use of
34
03SUT31836(2)_Color.indd 34
7/12/2019 6:57:51 AM
Underwater Technology Vol. 36, No. 2, 2019
more complex models or configurations. Further work is currently ongoing to prepare an AUV for a series of experiments to train its GPbased, nonparametric model. Once trained and verified, the vehicle will be deployed to assess the performance of the GPUKF outside a controlled environment.
References Allotta B, Caiti A, Chisci L, Costanzi R, Di Corato F, Fantacci C, Fenucci D, Meli E and Ridolfi A. (2016a). An unscented Kalman filter based navigation algorithm for autonomous underwater vehicles. Mechatronics 39: 185–195. Allotta B, Caiti A, Costanzi R, Fanelli F, Fenucci D, Meli E and Ridolfi A. (2016b). A new AUV navigation system exploiting unscented Kalman filter. Ocean Engineering 113: 121–132. Armstrong B, Wolbrecht E and Edwards DB. (2010). AUV navigation in the presence of a magnetic disturbance with an extended Kalman filter. In: Proceedings of IEEE Oceans 2010 Sydney, 24–27 May, Sydney, Australia. Bishop RED and Parkinson AG. (1970). On the planar motion mechanism used in ship model testing. Philosophical Transactions of the Royal Society of London 266: 35–61. Deisenroth MP, Huber MF and Hanebeck UD. (2009). Analytic momentbased gaussian process filtering. In: Bottou L and Littman M. (eds.) Proceedings of the ICML 2009 26th Annual International Conference on Machine Learning, 14–18 June, Montreal, Quebec, Canada, 225–232. Dunlap G and Shufeldt HH. (1969). Dutton’s navigation and piloting, 12th edition. Annapolis, Maryland: Naval Institute Press, 715pp. Ebden M. (2008). Gaussian processes for regression: a quick introduction. Robotics Research Group, Department of Engineering Science, University of Oxford. Available at <https://arxiv.org/pdf/1505.02965.pdf, last accessed 1 June 2019>. Fossen TI. (1994). Guidance and control of ocean vehicles. New York: Wiley, 480pp. Gertler M and Hagen GR. (1967). Standard equations of motion for submarine simulation. Bethesda, Maryland: David W Taylor Naval Ship Research and Development Center. Available at <https://apps.dtic.mil/dtic/tr/fulltext/u2/653861.pdf, last accessed 1 June 2019>. Hall R and Anstee S. (2011). Trim calculation methods for a dynamical model of the REMUS 100 autonomous underwater vehicle. Australian Government Department of Defence, Maritime Operations Division, Defence Science and Technology Organisation, DSTOTR2576. Available at <https://apps.dtic.mil/dtic/tr/fulltext/u2/a554483. pdf, last accessed 1 June 2019>. Kalman RE. (1960). A new approach to linear filtering and prediction problems. ASME–Journal of Basic Engineering 82: 35–45. Ko J, Kleint DJ, Fox D and Haehnel D. (2007). GPUKF: Unscented Kalman filters with Gaussian process prediction and observation models. In: Proceedings of the 2007 IEEE/RSJ International Conference on Intelligent
Robots and Systems, 29 October–2 November, San Diego, California, USA. Kuritsky MM and Goldstein MS. (1990). Inertial navigation. In: Cox IJ and Wilfong GT. (eds.) Autonomous robot vehicles. New York: SpringerVerlag, 96–116. Loebis D, Sutton R and Chudley J. (2003). A fuzzy Kalman filter for accurate navigation of an autonomous underwater vehicle. In: (eds.) Roberts GN, Sutton R and Allen R. Proceedings of the IFAC Workshop on Guidance and Control of Underwater Vehicles 2003. 9–11 April 2003, Newport, South Wales, UK 36: 157–162. Mandel J. (2009). A brief tutorial on the ensemble Kalman filter. Available at <https://arxiv.org/abs/0901.3725, last accessed 1 June 2019>. Ngatini T, Apriliani E and Nurhadi H. (2017). Ensemble and fuzzy Kalman filter for position estimation of an autonomous underwater vehicle based on dynamical system of AUV motion. Expert Systems with Applications 68: 29–35. Prestero T. (2001). Verification of a sixdegree of freedom simulation model for the REMUS autonomous underwater vehicle. BS thesis, University of California at Davis, Davis, California, USA. Available at <https://core.ac.uk/download/ pdf/4429735.pdf, last accessed 1 June 2019>. Rasmussen CE. (2004). Gaussian processes in machine learning. In: Bousquet O, von Luxburg U and Rätsch G. (eds.) Advanced lectures on machine learning. Berlin: SpringerVerlag, 63–71. Sen D. (2000). A study on sensitivity of manoeuvrability performance on the hydrodynamic coefficients for submerged bodies. Journal of Ship Research 45: 186–196. Telles da Silva Vale R, Apolonio de Barros E and de Castro Martins T. (2015). GPUaccelerated Monte Carlo localization for underwater robots. IFACPapersOnLine 48: 76–81. Turner R and Rasmusbluesen CE. (2010). Model based learning of sigma points in unscented Kalman filtering. In: Kaski S, Miller DJ, Oja E and Honkela A. (eds.) Proceedings of the 2010 IEEE International Workshop on Machine Learning for Signal Processing (MLSP 2010). 29 August–1 September 1, Kittilä, Finland, 178–183. Uhlmann JK. (1995). Dynamic map building and localization: new theoretical foundations. PhD thesis, University of Oxford, Oxford, UK. Available at <http://faculty.missouri.edu/uhlmannj/Dissertationpref.pdf, last accessed 1 June 2019>. Vio RP, Cristi R and Smith KB. (2016). Near realtime improved uuv positioning through channel estimation – the unscented Kalman filter approach. In: Proceedings of MTS/IEEE Oceans 2016 Monterey, 19–23 September, Monterey, California, USA. Williams CKI and Rasmussen CE. (2006). Gaussian processes for machine learning. Cambridge: The MIT Press, 248pp. Xiang X, Yu C and Zhang Q. (2017). Robust fuzzy 3D path following for autonomous underwater vehicle subject to uncertainties. Computers & Operations Research 84: 165–177. Zhang H, Xu Y. and Cai H. (2010). Using CFD software to calculate hydrodynamic coefficients. Journal of Marine Science and Application 9: 149–155.
35
03SUT31836(2)_Color.indd 35
7/12/2019 6:57:51 AM
Society for Underwater Technology
Educational Support Fund Sponsorship for Gifted Students in Marine Science, Technology and Engineering to meet industry’s critical shortage of suitably qualiﬁed entrants.
SUT sponsors UK and overseas students (studying in the UK and abroad) at undergraduate and MSc level who have an interest in marine science, technology and engineering. Students are supported who are studying subjects such as:
Offshore and Ocean Technology Subsea Engineering Oceanography Marine Biology Ship Science and Naval Architecture Meteorology and Oceanography The SUT annual awards are £2,000 per annum for an undergraduate, and £4,000 for a oneyear postgraduate course. (Parttime postgraduate studies funding available.) As one of the largest nongovernmental sources of sponsorship, the SUT has donated grants totaling almost half a million pounds to over 270 students since the launch of the fund in 1990.
For further information please contact Society for Underwater Technology, Unit LG7, 1 Quality Court, London WC2A 1HR UK t +44 (0)20 3440 5535 e info@sut.org or please visit our website
www.sut.org 03SUT31836(2)_Color.indd 36
7/12/2019 6:57:51 AM
Project Azorian: The CIA and the Raising of the K129 Norman Polmar and Michael White Published by Naval Institute Press
Hardback edition, 2010 ISBN 9781591146902 276 pages Although this book was first published in 2010, the story it tells and the descriptions of the technology used make such an extraordinary tale that it should be essential reading for anyone involved in underwater technology – especially for those with an interest in submarines, deepwater salvage, history, espionage, mining and even policy development. This is the tale of how the Central Intelligence Agency (CIA) attempted to salvage the remains of the sunken Soviet ballistic missile submarine, K129, from over 5000 m depth in the North Pacific Ocean in 1974. The book describes how this was done under a shroud of secrecy, with the help of billionaire Howard Hughes, by means of a cover story that suggested the lift ship, Hughes Glomar Explorer, specially built for the operation, was conducting trials for deepsea mining. The book is written by Norman Polmar, Secretary of the US Navy’s Research Advisory Committee at the time of writing, and Michael White, a TV and film producer who went on
to make the documentary, Azorian: The Raising of the K129. Notably, the authors have been required to omit some details of the operation from the book, owing to the level of classification that still applies to the project. In the acknowledgments, they list over 30 technical experts, military personnel and participants in the 1974 recovery operation, including individuals who were serving in the Soviet Navy at the time of the sinking of K129. The story is told in a documentarylike style, as might be expected given that Michael White was also producing a TV programme on the subject at the time. As one reads it, there are tantalising hints that the authors, or their sources, know a lot more than they are allowed to tell the reader – but even what can be told is a story that makes most spy stories seem mundane. K129 was a Project 629A (NATO ‘Golf II’), diesel powered, ballistic missilecarrying submarine. The class carried three largeyield nuclear missiles in launch tubes positioned in an elongated fin structure, and in the period before more capable nuclearpowered vessels were available, provided an important capability for the Soviet Forces. K129 sailed from its base on the Kamchatka Peninsula on 24th February 1968 to carry out a patrol within missilelaunching range of the US bases on Hawaii. After failing to receive a planned transmission by 8th March 1968, the Soviets commenced a search for the missing submarine on 10th March 1968. The sudden exodus of 36 warships and auxiliaries, numerous aircraft and at least five submarines from far
www.sut.org
Book Review
doi:10.3723/ut.36.037 Underwater Technology, Vol. 36, No. 2, pp. 37–39, 2019
eastern bases into the Pacific quickly attracted the attention of the Americans, who understandably wanted to know what was going on. After 72 days of searching, the operation was called off on 5th May 1968. Ninetyeight men were missing at sea and presumed lost, and the Soviets seemed to have no idea where the wreckage was actually located. Such was the secrecy culture of the former Soviet Union that the list of casualties wasn’t even published until November 1993, 25 years after the submarine’s loss. The Soviets had deployed acoustic arrays in the Pacific, but none had picked up the noises indicative of the loss of a submarine. Meanwhile, the US Navy cable ship, Albert J Myer, was carrying out acoustic surveys for a planned defence acoustic array site in the Eastern Pacific, about 1700 nautical miles from K129, and on 11th March 1968 had detected two ‘acoustic events’. At that point, the US Navy’s Deep Submergence Systems Project office, and its chief scientist Dr John Craven, became involved and found other acoustic records from arrays that had been established to detect clandestine underwater nuclear tests in the Pacific. Analysis of five acoustic data sources pinpointed a position to within two nautical miles of 40 degrees 6 minutes north, by 179 degrees 57 minutes west, occurring within 1 second of 1200 Zulu on March 11th 1968. This position was about 600 nautical miles from where the Soviets had focussed their search for K129. Interestingly, these acoustic events occurred a few days after K129 failed to report in, fuelling speculation over the
37
04SUT32736(2)_Color.indd 37
7/12/2019 7:01:09 AM
Polmar and White. Project Azorian: The CIA and the Raising of the K129
years about what the submarine was up to. The wreckage was estimated to be lying at a depth of 16 800 ft – within range of US Navy assets such as the bathyscaph, Trieste, or systems towed by the research ship, Mizar. However, deploying these systems would have given away that the Americans were looking for a deep object and alerted the Soviets that K129 had been found. Fortunately, the nuclear submarine, USS Halibut, had been secretly converted by Dr Craven’s team to take part in the ‘Sand Dollar’ programme, which involved using early remotely operated vehicles (ROVs) and towed camera systems to search the sea floor for useful debris, such as reentry vehicles and warhead components, from Soviet missile tests. Halibut could explore the sea floor for extended periods deploying an array of systems – for any member of SUT these capabilities being available as early as the late 1960s is quite an eye opener! The costs of refitting Halibut and developing lights, cameras and sonar capable of surviving at 20 000 ft was huge; these costs were hidden inside ‘cost overruns’ on the Polaris and Poseidon nuclear programmes – a tactic used many times over for hiding the costs of ‘black projects’ by governments. Halibut started the search for K129 in April and May 1968, deploying navigational transponders through torpedo tubes; after more than 20 000 photographs were taken of the area, it found the wreckage of K129 in two pieces. Analysis showed that one of the missile tubes was intact, along with the key compartments likely to contain code books, documentation, cryptographic machinery and other treasures that would be very worthwhile to recover and examine in detail. The American
intelligence community wanted to know if a nuclear missile and the other components could be recovered from such depth, without the Soviets realising what was going on. The CIA and the US Secretary of State, Henry Kissinger, became involved, and eventually President Nixon signed off a secret project to attempt to recover the forward section of the submarine. A team led by John Parongosky examined various methods to lift the wreck, with at least three concepts examined in detail: brute force lifting by winch; a threemile drillstring with claws; and two approaches to generate excess buoyancy. In November 1969, CIA officials met with the vice president of engineering from Global Marine of Los Angeles, and requested the company to build something that could lift the number of tons required from a water depth of 15 000–20 000 ft. Global Marine had been chosen because of its expertise in dynamic positioning and experience gained building the drill ship, Glomar Challenger. A team was assembled, geophysics were completed on the deepocean floor in the vicinity of K129 to ascertain the ground conditions, and construction then began on the Glomar Explorer to recover and transport the stricken submarine. Interestingly, Soviet intelligence did get a sniff that the Americans were up to something, but it dismissed the idea as technically impossible at the time, and only later realised what had been going on. As it was hard to hide a massive ship under construction and the associated materials, the book explains how in late 1971 someone came up with idea of using Howard Hughes as part of a cover story for the whole incredible operation. The cover story feigned that the eccentric
billionaire was building a ship to explore for and recover manganese nodules, a vast untapped resource on the sea floor. Indeed, the cover story was so effective that well into the 1980s university textbooks showed photos of the Glomar Explorer alongside photos of manganese nodules. The government of Malta became concerned that the United States would start hoovering up seabed resources, the ‘common heritage of mankind’, in international waters, and their concern directly influenced United Nations (UN) policy, the UN Convention on the Law of the Sea, and the creation of the International Seabed Authority – an unintended consequence of a CIA cover story! The new ship for the recovery required a vast gimballed platform to isolate the suspended load (K129) from pitch and roll; heave compensation; hydraulic lift; pipehandling; a moonpool large enough to accommodate a submarine’s forward section; a docking system for the ‘capture vehicle’; and dynamic positioning. The ship, christened the Hughes Glomar Explorer, was launched on 4th November 1972 and repositioned to Long Beach, California to have the additional ‘black ops’ features installed. These included 24 containerised offices comprising dark rooms, radioactive decontamination equipment, sample preserving facilities, and so on. Pipestrings were fitted to enable operation at 17 000 feet, and a floating hangar known as the Hughes Mining Barge 1 was built to house K129. A capture vehicle nicknamed ‘Clementine’ held the claws required to grasp the submarine’s hull and raise it off the seabed. Staying true to the seabed mining cover story, curious onlookers would be told that the secret nodule harvesting
38
04SUT32736(2)_Color.indd 38
7/12/2019 7:01:10 AM
Underwater Technology Vol. 36, No. 2, 2019
equipment was housed inside the ‘mining barge’ that would store the precious cargo. Time was moving on, and in Spring 1974 Henry Kissinger signed off the evaluation that even after 6 years, the intelligence dividend of recovering the nuclear warhead and other materials, especially cryptographic equipment, still justified the huge expenses incurred. The Hughes Glomar Explorer then set sail on 21st June 1974, with an ‘A’ crew cleared for the initial recovery phase and trained to recognise Cyrillic alphabet signs such as ‘danger’ and ‘radioactive’, and a ‘B’ crew tasked with transferring material to storage, hiding unwanted remains (presumably back on the sea floor?) and ‘cleaning up’. The authors cover in some detail the actual recovery operation, complete with names of key personnel; they explain that the operation was only partly successful, as part of the hull of K129 fell away during recovery, taking the missile and cryptographic equipment with it. However, the nucleartipped torpedoes were recovered, together with human remains in good condition. These were given a suitable funeral at sea, and after the Cold War ended a video tape showing the funeral and the respect shown towards the bodies, was passed to President Yeltsin. Mechanical failures in the lifting mechanism were blamed for the loss of the rest of the hull, and although plans were made to return and finish the
job, eventual Soviet discovery of the operation (likely via spy activities) rendered it impossible to complete without incurring major diplomatic or even military resistance. As such, much of K129 still rests deep in the darkness of the North Pacific Ocean. The story ends with an analysis of whether the operation was worth it, arguing that although (as far as the authors can tell us…) the project failed in terms of its primary goal to recover the missile and cryptographic materials, it was successful in that it did recover other significant material such as nucleartipped torpedoes. The project also taught the US naval and subsea engineering community valuable subsea engineering and technology skills that went on to have tremendous value to offshore hydrocarbons, defence and other sectors. The operation also drove the development of deepcapability systems capable of recovering objects from the seabed in a far more clandestine manner than Project Azorian could do. In telling the story, the book describes the use of underwater passive hydrophone arrays that enabled the discovery of the wreck, showing that even in 1968 it was possible to detect a submarine at a range in excess of 3000 nautical miles in optimum conditions. It covers the evolution of the early submarinelaunched ballistic missiles, with their troublesome liquid fuels, and the SSN5 ‘Serb’ missile carried by
K129. Although by the late 1960s the Golfclass submarines were nearing obsolescence, they still carried stateoftheart Soviet technology in their various systems, in addition to code books, cipher equipment, sealed orders. Of course, the class also contained nuclear warheads from the ‘Serb’ missiles and two nucleartipped torpedoes; the prize was worth the effort. The book contains a centre section of speciallycommissioned colour illustrations that give impressively detailed pictures of the equipment used. In fact, they are the most detailed that this reviewer has come across to date. The detailed appendices at the end of the book include sections on the likely causes of the loss of K129, some of the conspiracy theories, technical details of the USS Halibut and a full list of the crew who were lost onboard K129. It may be some years before all the details pertaining to Project Azorian – and indeed some other extraordinary tales from the Cold War subsea world – are fully known. But for now, this book is the closest we can get to discovering the full story of K129, and the boost the programme to recover the wreck gave to the ocean technology community, benefitting so many of our activities such as seabed mining that only now, some 50 years on, is becoming a reality. (Reviewed by Stephen Hall, Chief Executive, Society for Underwater Technology)
39
04SUT32736(2)_Color.indd 39
7/12/2019 7:01:10 AM
Society for Underwater Technology International multidisciplinary learned society This nonaligned membershipbased organisation seeks to further the dissemination of knowledge and lessons learned in the underwater environment through networking, events and publications
Its membership covers the following activity areas:
diving and manned submersibles environmental forces marine policy marine renewable energies ocean resources offshore site investigation and geotechnics salvage and decommissioning subsea engineering and operations
For further information For membership, publications or general enquires, contact SUT Head Ofﬁce Unit LG7, 1 Quality Court, London WC2A 1HR t +44 (0)20 3440 5535 e info@sut.org For events, contact SUT Aberdeen Ofﬁce Enerprise Centre Exploration Drive Bridge of Don Aberdeen AB23 8GX UK t +44 (0)1224 823 637 e events@sut.org
underwater robotics underwater science
www.sut.org
underwater vehicles
04SUT32736(2)_Color.indd 40
7/12/2019 7:01:11 AM
UT2 and UT3 The magazines of the Society for Underwater Technology
UT2 Issue O
UT2 Issue O
ne 2018
A Towed Vehic les CaSubsea Expo Roundup Un
1
A Ca Un THE
MAG
AZIN
E OF THE
SOCI
ETY
FOR
UNDE
UT2 Iss ue 1 Ma rch 2018
Decem
ber 20 15 Janu ary
ne 2016
OSIG Vess el Batteries Underwat er Vehicles
RWAT
1
ue 2OAp LOGY ril June 2017
Issue Tw
The magazine is represented at all the many exhibitions around the world at which the Society both coorganises and attends.
EUT R TE CHN 2 Iss
UT 2016
UT2 covers a focused range of underwater subjects including offshore, marine renewables, subsea engineering, ocean resources, diving and manned submersibles, underwater science and robotics.
Furthermore, the magazine is distributed at the many subsea training courses that are organised by the Society, ensuring it reaches tomorrowâ&#x20AC;&#x2122;s engineers and technologists.
o 2018
UT3 is the online magazine of the Society for Underwater Technology, and covers the subsea industry.
A Ca Un THE
MAG
AZIN
E OF THE
SUT OCI
1
ETY F 3 Au g OR U NDER Sept 2015W A
Subsea Mi ning Underwat er Vehicles
TER
TECH
NOLO
GY
AUVs Cables Unmanne d
Vehicle
1
UT2 Iss ue 2 Ma y 2018
It consists of the content of the print magazine UT2, greatly expanded with other information.
UT22 and UT33 are available online at http://issuu.com/ut http://issuu.com/ut2_publication 2_publication www.sut.org 05SUT36(2)IBC.indd 1
7/12/2019 6:41:43 AM
BOOK YOUR STAND TODAY
CALLING ALL OCEAN TECHNOLOGY SUPPLIERS The Global Meeting Place for the Ocean Technology Community Oceanology International will connect you with 8,000 buyers and influencers from more than 70 countries looking for innovative solutions to improve strategies for exploring, monitoring, developing and protecting the worldâ&#x20AC;&#x2122;s oceans.
MEET different geographies in one location at one time
REACH multiple buyers from multiple industries all under one roof
CONNECT with new and existing clients facetoface
DISCOVER cutting edge content that will attract a highlevel audience seeking the latest developments
Enquire to exhibit at oievent.com/enquire Organised by:
06SUT36(2)OBC.indd 1
In partnership with:
Endorsing associations:
7/12/2019 6:43:06 AM