69
1 INTRODUCTION
The Automatic Identification System (AIS) is an
automatic tracking system for identification and
location of vessels by exchanging data via VHF
communication to other nearby ships, AIS base
stations, and satellites. It has become mandatory,
through the International Maritime Organization
1
(IMO),forcommercialvesselsover300GTsince2004
affecting approximately 100.000 vessels. Additional
legislation from the EU and the US extended the
requirementsforhavinganAIStransmitteronboard
also to smaller crafts such as fishing boats. It is
1
http://www.imo.org/en/OurWork/Safety/Navigation/Pages/
AIS.aspx
estimated that in 2012 there were a quarter of a
million vessels equipped with AIS and that this
numberwillrisetooveramillioninthenearfuture,
asperWikipedia (2015).SinceAIS transponders use
VHFcommunicationthereliablerangeisabout10‐20
nauticalmiles,althoughAISsatellites
canpickupAIS
signalsfromspace.TheoriginalpurposeofAIS was
meantasanaidtocollisionavoidancebutmanyother
applications have since been developed such as
fishingfleetmonitoring,vesseltrafficservices(VTS),
maritime security, fleet and cargo tracking, search
and rescue, accident investigation among others.
In
addition to the increase in the number of AIS
transponders, there is an extensive effort, both
commercially and by governments, to increase; the
global coverage of the AIS signal and the volume
Identifying and Analyzing Safety Critical Maneuvers
from High Resolution AIS Data
T.Mestl,K.T.Tallakstad&R.Castberg
DNV‐GL,ResearchandInnovation‐ITAnalytics,Høvik,Norway
ABSTRACT:WedemonstratethevalueinpreviouslydisregardedparametersinAISdata,andpresentanovel
wayofquicklyidentifyingandcharacterizingpotentiallysafetycriticalsituationsforvesselswithaproperly
configuredAIStransponder.Thetraditionalapproachofstudying(near)collisionsituations,
isthroughvessel
conflictzones,basedonvessellocationandspeedfromlowresolutionAISdata.Ourapproachutilizestherate
ofturnparameterintheAISsignal,atmaximumtimeresolution.Fromcollisioninvestigationreportsitisoften
seenthatpriortooratcollisionnavigatorsperformfreneticrudder
actionsinthehopetoavoidcollisioninthe
last second. These hard maneuverings are easily spotted as non‐normal rate of turn signals. An identified
potentialcriticalsituationmaythenbefurthercharacterizedbytheoccurringcentripetalaccelerationavesselis
exposedto.Wedemonstratethenoveltyofourmethodology
inacasestudyofarealshipcollision.Astherate
ofturnparameterisdirectlylinkabletothenavigatorbehavioritprovidesinformationaboutwhenandtowhat
degree actions were taken. We believe our work will therefore inspire new research on safety and human
factorsasa
riskprofilescouldbederivedbasedonAISdata.
http://www.transnav.eu
the International Journal
on Marine Navigation
and Safety of Sea Transportation
Volume 10
Number 1
March 2016
DOI:10.12716/1001.10.01.07
70
receiving capabilities of satellites, base stations, and
data stores
2
. This follows the Big Data trend, and
additional insights may be gained from high time
resolutionAISdata.
AIS is an important source of information for
studying maritime traffic and associated critical
situations, in particular ship‐to‐ship collisions. Most
studiesonriskofship‐to‐shipcollisionsarebased
on
identificationofpotentiallycriticalcollisionsituations
intheAISdata.Interestingly,neitherIMOnorcourts
state explicit criteria defining critical situations or
collisionrisks.Theclosestdefinitionofa nearmissby
IMOis”a sequenceofeventsand/or conditions that
couldhaveresulted inloss.Thislosswas
prevented
only by a fortuitous break in the chain of events
and/orconditions.”(IMO(2008)).According toSturt
(1991),”...thatmustalways bedecided,accordingto
the circumstances of each case, by men of nautical
experience”. This means the interpretation of near‐
collisionsisalmosttotallysubjectiveasitwill
depend
on the ”comfort zone” of the involved parties. One
mastermayconsideracertaindistanceassafeenough
whereas others may not. Also, the acceptable
minimum distance between vessels will most
certainly depend on the size of the vessels, their
(relative) speed, their maneuverability, maybe their
cargo and of
course on a number of occurring
circumstances such as sea state and weather
conditions. All these factors will influence what is
considered as ”close quarters” or the size of the
comfortzone.
Considerableefforthasbeenputintoderivingan
operational concept that defines a domain around a
shipthatwould
constitutea”vesselconflictzone”,i.e.
a geometric area where there is a probability of
collision. The sizes and shapes of theses domains
range from very simple to quite complex structures,
some are circular, elliptical or polygonal, some are
static,andothersaredynamicallyresizeddepending
onthespeedof
thevessel,e.g.Fujii&Tanaka(1971),
Goodwin (1975), Coldwell (1983), Zhao etal. (1993),
Pedersen (1995), Mestl et al. (2008), Pietrzykowski
(2008), Pietrzykowski & Uriasz (2009), Wang et al.
(2009),Zhangetal.(2015)inordertomaketheterm
”near collision” and ”comfort zone” more tangible
andquantifiable.
So
far the available approaches for identification
and quantification of critical situations are either in
form of a subjective zone or on a numeric ”fear
factor” as per IMO (2015), expressing the perceived
risk of a close encounter. The validity of the zone‐
approach has already been questioned in the
literature
as the intended vessel movement,
background traffic, ship type and hydro‐
meteorological conditions should also be taken into
account when determining whether a dangerous
situationoccursornot,Goerlandtetal.(2012),Kaoet
al. (2007), Montewka et al. (2011). It may also be
pointed out that all the traditional approaches
are
quite complex, not only in the construction of the
zone and fear factor, but they are also quite
computational intensive when trying to identify
potential crossing of ship trajectories. To keep the
2
See e.g. companies such as VesselTracker, Orb-
Comm, exactEarth, Spire etc.
computational work load manageable all these
approaches utilize down‐sampled AIS data, i.e.
usually 6‐10 minutes time difference between
samples, and they only use the geo‐locations of the
vessel.OtherrelevantparametersofferedbytheAIS
signalaretoalargeextentneglected.
Inthispaperwewill
presentanewapproachthat
allows identifying potential critical situations rather
quickly, in historical time series of AIS data. This
methodutilizeshighresolutionAISpositiondata,and
the rate of turn parameter available in the data
stream.Thenextchapterpresentsacasestudyofan
actual ship collision found
in our data material,
demonstrating the advantage of rate of turn (ROT).
We derive a general outline how to identify non‐
normal maneuvering and their characterization. The
paperisconcludedwithsomecriticalremarksabout
our approach, and open a discussion regarding
potentially necessary steps the IMO or other
authorities
maywanttotaketoincreasethebenefits
offeredbyAISdata.
2 METHODOLOGY
2.1 Identificationandcharacterizationof(near)collisions
bynavigationalparameters
Inthefollowing wewilloutline anewapproach for
identification and characterization of (near) collision
situationsthatutilizeshigh time frequency AIS data
(2‐12 second
sampling rate) using the following
parameters: latitude, longitude, rate of turn (ROT),
speed over ground (SOG), and course over ground
(COG). Note that the ROT does not represent the
rudderangleperunittime,buttheactuallyoccurring
change in heading of the vessel, as per IMO (2003).
Wehave
chosentopresentthedetailedfindingsofa
single incident to demonstrate the feasibility of our
method. It is left to future studies to estimate near
collisionfrequenciesinNorwegianwaters.
The following excerpt from a near‐collision
incidentreportunderlineswhyweconsidertheROT
as of one of the
most interesting AIS parameters. It
wasissuedbytheTheTransportationSafetyBoardof
Canada (1998), describing the circumstances around
the near miss between the cruise ship
”STATENDAM” and the tug/barge ”BELLEISLE
SOUND”/”RADIUM 622” in the Discovery Passage,
British Columbia on 11 August 1996: ”The hard‐a‐
starboard maneuver caused
the ”STATENDAM” to
heelovertoport,andresultedinsomeminorinjuries
to six passengers and two of the crew. ...”. Our
approach is based on the observation that any near
collision or actual collision is usually accompanied
with some frenetic activity right before the (near)
collision in the hope
to avoid it. Many critical
situations evolve because the navigators
assume/expect that the other party will do the
necessarymaneuversforcollisionavoidance.Evenif
the vessels can no longer avoid a collision the
navigator(s)willneverthelesstrytoturnthewheelin
the seconds before the impact in the
hope this will
diminishtheconsequences.Thismeans,insituations
wherethenavigatorhasrealizedthatasituationmay
become critical he/she will turn the wheel trying to
avoidthe criticality.In these situations relative high
71
ROT values should be observable in the AIS data.
Thus, we may anticipate that the closer a potential
collisionisintimeand/ordistance,themore intense
will be the navigational action of one or both
helmsmen. Due to the usually large inertia and
momentumofships,changingthespeed
ofavesselis
generally not considered an option resulting in fast
changes.Theshipsresponsetotherudderisusually
muchfaster(dependingonthespeed). According to
rule8inCOLREG‐PreventingCollisionsatSea(IMO
(1972))itisstatedthat”Actiontakentoavoidcollision
with another
vessel shall be such as to result in
passing at a safe distance.” No practically usable
information is given regarding the safe passing
distance except that ”it depends on the
circumstances” and that ”the person on the other
vesselshouldnotfeelcompelledtoactalsotoincrease
the distancefurther.”
Inthis respect we could claim
that our approach based on ROT is actually in
alignment with this fuzzy IMO requirements as it
actually focuses on the (rudder) action taken ”to
increase the distance further ” Llana & Wisneskey
(1991).
2.2 Closeupstudyofacollisionbetweenaferry
anda
fishingvessel
The NorwegianCoastal Authorities kindly provided
DNV GL with a high resolution AIS data set for
research purposes. We will therefore anonymize the
presented data as much as possible, i.e. remove
reference to the involved parties, time of occurrence
andgeo‐location.
The following collision occurred in
open waters,
far from a port, in Norwegian waters recently.
Fortunately there were no injuries or pollution, nor
were there any significant material damages.
According to the investigation report, on a summer
morning (08:45,localsummertime), a fishing vessel
in transit crossed the trajectory of a larger ferry (on
regularroute)fromaport.Thehelmsmanondutyon
theferrywasbusyoutsidethebridgeandthelookout,
not possessing a bridge certificate, alerted the
helmsman too late. According to the administrating
directoroftheferry line: ”...a number of maneuvers
wereperformed leadingtoa considerableheelingof
the vessel which is quite normal when using the
rudder a lot...”. This statement gives again an
indicationthattheremusthavebeenquiteahighrate
ofturnoftheferry.
Figure1showsthesevesseltraceswithalow(5‐10
min. sampling rate), i.e. traditional, and high
frequency (2‐10 sec. sampling rate) AIS data feed.
Basedonthelowresolutionsignal,itisindeedquite
difficult to determine whether a critical situation
occurred or not. Right before the incident, the ferry
hadaspeedof18knotsandanAISsamplewassent
on average every 3
seconds. For the fishing vessel,
travelingat8knots,theAISsamplingintervalwas11
seconds. The trace clearly indicates that the fishing
vessel did not show any evasive maneuvers before
the impact, whereas the ferry reacted too late. The
collision time can be extracted from the
latitude/longitudeparametersor
theCOG parameter
intheAISdatafeed.
Figure1. Visualization based on high resolution AIS data,
overlaid with the low resolution AIS data points (blue
triangle‐ferry,greentriangle‐fishingvessel).Notethatthe
ferry’slowresolutionAISpointatthe”normalturnaround’
couldbemistakenasanoutlier.Thehighresolutiontraceof
the ferry is
colorcodedaccording to its rate ofturn (blue
lowROT,redhighROT).Thecollisionpointsintheirtraces
areindicatedwithcyandots.
ThetoppanelinFigure2 showstheCOGforboth
vessels.
Note that the fishing vessel shows no sign of
evasivemaneuversuntilcollisionwhichappearsasa
sudden change in COG, coinciding with that of the
ferry. We can therefore deduce that the collision
occurred between 7:23:00 and 7:23:07
(UCT) on that
morning. The AIS transmitter’s sampling rate
dependsonvelocityandturnrate,withamaximum
of2samplespr.secondforhighvelocityorturnrate.
Thus, due to the relatively low speed of the fishing
vessel,the correspondingAISsamplingintervalwas
notatmaximum,hence
the∼7secondsuncertaintyin
thecollisiontime.Observealsothatthefishingvessel
(∼300GRT)felttheimpactmuchmorethantheferry
(∼5700GRT),hencethenoticeablechangeinCOG.
ThebottompanelinFigure2showstheROTofthe
ferry.Unfortunatelythefishing
vesselwasnotsetup
tologROTvalues.Noticethe highpeakin theROT
indicating a hard starboard (positive values)
maneuver right before the collision revealing the
futileattemptoftheferrynavigatortoavoidcollision.
Inthe7secondstimeuncertaintyofcollisiontime,the
ferry sent three
AIS samples due to her higher AIS
sampling rate (shown as cyan dots). It is not
surprisingthatthehighrateofturnwasfeltslightly
uncomfortable by some passengers, as the comfort
limit for cruise ships (at 20 knots) is considered 10
deg.pr.mininROT,asperThe
TransportationSafety
BoardofCanada(1998).Theevasivemaneuverofthe
ferry(at18knots)wasfarabovethatlimit.Inorderto
relate the trace and COG to the occurring ROT, the
corresponding trace (see Figure 1 and COG (see
Figure2bottompanel)oftheferrywascolor
coded
basedonherROT(maximummeasuredROT=194.5
deg/minandROT=0.0deg/minarecolorcodedred
andbluerespectively).
72
Figure2.TOP‐Courseoverground(COG)forbothvessels
(green squares‐fishing vessel, blue/red dots‐ferry).
Observe that the COG for the fishing vessel remained
unchanged until collision, after which it changes abruptly
and parallels with that of the ferry (also seen on the geo
tracesinFigure1).From
thechangesinCOG,thecollision
must have occurred between 7:23:00 and 7:23:07 (UCT) as
indicatedwiththecyandash‐dottedlines.BOTTOM‐Rate
of turn (ROT) of the ferry in deg/min. Observe the high
peak (hardstarboard maneuver) right beforethe collision.
Theredandbluecolorbarrepresents
highandlowvalues
ofROTrespectively.Forcomparison,thegreendashedline
representstheROTcomfortlimitforcruiseshipstraveling
at 20 knots. Thus, it may not be surprising that the ferry
maneuvers were experienced asslightly uncomfortable by
passengers.
Figure3.LEFT‐Thedailymaximum|ROT|values(upper
reddots),thedailymedian |ROT|values (bluedots),and
the |ROT| values corresponding to the 95 % percentile
(orangedots) ofthe ferryoverhalf ayear. OnlyROT /=0
were taken into account. The highest ROT peak relates to
thecollisionwiththefishingvessel.Therelativehighpeak
thedaybeforecorrespondstoasharpmaneuveringduring
thepassagebetweenagroupofislands(seeFigure4).Itis
easy to pick out any non‐normal maneuvering from this
singleshiptimeseriesplot.Hereall|ROT| ≥
150deg/min
aretaggedwiththedateofoccurrence.RIGHT‐Frequency
ofoccurrenceofvarious|ROT|values.Notethat99.999%
ofall|ROT| ≤ 150deg/min.
2.3 Generalapproachforidentificationand
characterizationof(near)collisions
ThecaseexampleinSection2.2clearlydemonstrated
thevalueoftheROTparameterintheAISdatafeed.
In the following we show how potentially critical
situationscanbeidentifiedquickly.Wethenturnour
attentiontocharacterizingthesesituations.
Figure 3 showsthe various daily maximum ROT
valuesoftheferryoverhalfayear,closeto2million
samples for the full dataset. Each vertical line
represents one day. In Figure 3, it is seen that for
somedays theferry stayedinport,givingonlyzero
values,
whereas no samples were available over a
numberofdaysperiodinApril.Thehighestpeakin
ROT relates to the collision of the ferry with the
fishing vessel. Interestingly, there is also a relative
high peak the day before which after a closer
examinationturnsouttohavebeena
verysharpand
definitively a non‐normal turn when navigating
throughagroupofislands.ThisisshowninFigure4,
wherethetrackhasbeencolorcodedaccordingtothe
ROT. Notice the very high values along the sharp
turn,comparedtomorenormalbehavioronprevious
passages. This
might have been a situation where a
non‐AIS emitting object was encountered, since no
other vessels where present nearby. The bar plot in
therightpanelinFigure3isahistogramshowingthe
frequency the various ROT’s occurring over half a
year.AROTvalueabove150deg/minis
veryrare,i.e.
99.999% ofallsampleswerebelow this value.Note
that this is not the fraction of safe passages, but the
number of ROT signals above the 99.999 % quantile
for that vessel. A high ROT can therefore be
considered as an indication of a potentially non‐
normalmaneuvering.Thetime ofitsoccurrence can
beobtainedfromthetimeseriesplot.Ameasurefor
sensitivity of detection of non‐normal navigational
maneuversbasedonROTisgivenbyahighsignalto
noiseratio.By usingmax(|ROT|)/median(|ROT|)∼
36, we see that the collision incident
was far above
whatisexpectedasanormalROTfluctuation.There
were four other occasions where the |ROT | > 150
deg/min, i.e. on 20th February, 20th March, 16th of
Apriland15thof May.A closerinspectionindicates
thatmostofthesehighROT’sweresingleoccurrences
which
may be regarded as outliers, see for example
Figure8inSection3.2.
Large ROT’s are necessary but not sufficient
indicators of potential evasive maneuvers. A further
characterization of an identified high ROT is its
associatedcentripetal acceleration(CA),i.e.changein
rotationaroundthe vertical axis.Thisaccelerationis
of
interest as passengers will be exposed to it when
the ship is turni ng. The magnitude of centripetal
accelerationisdefinedas
2
SOG
CA
R
(1)
whereRistheturnradius.Thetime Tittakesfora
vessel with speed SOG to complete a whole circle
withradiusRis
2
R
T
SOG
(2)
73
which must be the same time it takes for the vessel
withaconstantROTtocompleteawholecircle,i.e.
2
T
R
OT
(3)
Equaling both expressions of T , solving with
respect to R, and inserting it into Eq. (1) we get an
expression for the centripetal acceleration (CA) in
termsofROTandvesselvelocitySOG:
2
SOG
CA SOG ROT
R
(4)
Figure4.AIStrackofaferrynavigatingthroughagroupof
smallislands.Thetrackhasbeencolorcodedaccordingto
therateofturn(blue‐lowROT,red‐highROT).Notethe
verylargeROTvaluesofthelowermosttrackcomparedto
the more normal values on previous passages
around the
islands.
The advantage of using the CA is that it directly
relates to the human perception of comfort. A daily
maximumplotofCAvaluesfortheferryareshown
in the left panel in Figure 5. The stated maximum
ROT= 10deg/minasthecomfortlimitoncruiseships
at
20knots,asperTheTransportationSafetyBoardof
Canada (1998), translates to CA = 0.03 m/s
2
, and is
shown by the green dashed line in the left panel. It
maybequestionediftheCAcomfortlimitofacruise
ship is directly transferable to a relative small ferry
navigatinginharshseas.Onsuchaferry,passengers
may have to expect larger vessel motion,
i.e. rolling
and stamping but also higher turn accelerations. In
contrast to the ROT, which directly measures an
evasive maneuver, the CA may be considered a
measure of passenger comfort. From the CA time
seriesinFigure5,we cansee thatinadditiontothe
previously identified occurrences of high
ROT’s we
havehighCAsonanumberofotherdaysaswell,e.g.
9th January, 27th March, 10th April, 9th May and
22nd June. Due to a lower speed over ground, the
occurrenceofahigh|ROT|on15thofMay(seenin
Figure3)doesnotcoincidewith
acorrespondinghigh
|CA|atthesamedate.Definingasignaltonoiseratio
similar to what was done to |ROT|, i.e.
max(|CA|)/median(|CA|) ∼ 40, we see that the
centripetal acceleration gives a slightly better
detectionperformance.Thebarplotintherightpanel
inFigure5isahistogramshowing
thefrequencythe
various |CA|’s recorded over half a year. A |CA|
valueabove0.35deg/minisveryrare,i.e.99.999%of
allsampleswerebelowthisvalue.
The degree of discomfort a passenger may
experience,willnotonlydependonthemagnitudeof
acceleration,butalsoonhow
longitwilllast.Ahigh
accelerationoverjustafewsecondsmaybefeltasa
sideways bump and may rather be considered as
annoying asonemightspillthedrink.Ontheother
hand,ahighcentripetalaccelerationoveranextended
periodoftime,couldresultinconsiderableheeling
of
thevessel.ByintegratingallCApeaksintime,above
a given threshold, one obtains a useful metric for
passenger/crew discomfort. This quantity is then
proportionaltotheimpulseexperiencedwhenaship
turns.Asweareonlyconcernedaboutsafetycritical
situations, which are characterized by excessive
maneuvers,
this threshold may be defined e.g. as
thosecaseswhereboththe|ROT|andthe|CA|are
abovetheir99%percentile.
Figure5.LEFT‐Timeseriesofdailymaximum|CA|over
half a year, only CA /= 0 were taken into account. The
comfortlimitforcruiseshipsisgivenbythegreendashed
line.SimilarlytoFigure3,max(|CA|),median(|CA|),and
95%quantilearecolorcodedbyred,blueand
orangedots
respectively. In the time series, some pronounced
accelerationpeaks(taggedwiththedateofoccurrence)are
different from those in Figure 3. Interestingly, the high
|ROT| on the 15thMay has a much lower corresponding
CA (vertical gray dashed lines indicates extreme values
from ROT in Figure 3).
RIGHT‐Histogram of the
acceleration values with the corresponding fraction of
samplesonthex‐axis.
This measure may be useful for providing a
general characteristic of how a captain handles the
ship.
3 RESULTSANDDISCUSSION
3.1 UnderlyingcausesofextremevaluesinROTand CA
Theresultofacloserinvestigationoftheemphasized
dates from Figures 3 and 5 is given in Table 1.
The
chosen dates correspond to occurrences of high
|ROT| and |CA| values, that is, values above the
corresponding 99.999 percentiles (|ROT| > 150
deg/min and |CA| > 0.35 m/s
2
). Note from Table 1
thatthecollisionsticksoutforbothROTandCA.A
highROTmaygiveafirstbutnotsufficientindication
ofapotentialnon‐normalmaneuvering.Forexample,
74
the high |ROT| on 20th of March and 15th of May
occurred during heavy sea maneuvering. It is
thereforedifficulttouseasinglemeasurereliably,in
order to identify and characterize non‐normal
maneuverings. However, both the ROT and CA
measures, possibly combined with the impulse
measure described in
the previous paragraph, will
provide valuable information about various
navigationalaspects.ThescanforlargeROTvaluesis
computationally cheap and effective, for a first
narrowdownoflargedata sets.Asanexample,the
datasetinourcasestudywasreducedfrom2million
pointstoonly6,
whichthencanbefurtheranalyzed.
For this particular ferry no other near collision
incidents could be identified neither with our
methodology,norwiththetraditionalzoneapproach
over the half year. This is not surprising, since the
navigators’taskistoavoid(near‐)collisions.
Our methodology flags all non
‐normal
maneuverings which may however not always
represent safety critical situations. As pointed out
previously, it is also important to realize that non‐
normalmaneuversmayoccurevenwhenthereareno
other vessels nearby. For instance, a floating
container,fishingnetsoraleisureboatmayforcethe
helmsman
to suddenly change course. Illustrative
examples are shown in Figure 6. It seems that this
ferry had to turnaround in a hurry, hence the high
ROT.Ifwetakeintoaccountthetimeofyear,i.e.early
spring in Norway (8th May), late evening (21:53
oclock),itscoastallocation(remotearea),
andthefact
that the vessel leaves the harboragain just after a 3
minutes stay, one may conclude that the ferry was
probablyemptyanditturnedaroundtopickupalate
car arrival. By taking into account the ferry liners
arrive and departure schedule, it turns out
that the
situation was rather critical for the car, as the last
ferry departure was scheduled for 21:45. Such a
turnaround is a common phenomenon in remote
Norwegiancoastalareas,andmayratherbeclassified
ascustomerservice.
3.2 ShipspecificdependenciesofROTandCA
Note that the maneuvering measures,
i.e. ROT and
CA,willbevessel specific(Bertram(2000);Rawson&
Tupper (2001)). A small ferry vessel is much more
maneuverable than a large tanker or freight ship.
Hence,therangeofROTandtherebyalsoCAwillbe
quitedifferent.Thetoppanel in Figure 7 shows the
different observed ROT values for a small ferry
comparedtoalargetanker,randomlyselectedinour
dataset.Bycomparing thedistributions, itisevident
thattheferryhasalargerspreadoftheROTvalues,
implyinghighermaneuverability.Notealsothatthea
ships ROT will not only depend on
ship
characteristics, but also on the sailing pattern, e.g.
navigationthrougharchipelagoversusopensea.
Consequently, one has to be cautious when
comparing these measures between vessels. Only
vessels of the same size, draught, maneuverability
may be compared. The bottom panel in Figure 7
shows the dependency between ship length
and the
averagemaximum|ROT|values,forallvesseltypes,
inourdataset. Roughly1000vessels areincludedin
thisstatisticalaverage,i.e. allvesselsinourdatabase
with a meaningful ROT signal. This dependency of
ROT on vessel characteristics makes it more
cumbersometoidentifyextraordinarymaneuverings,
as one has
to define, for each individual vessel (or
groups of similar vessels), ”normal” maneuvering
behavior.Thisalsomeansthatenoughsamplesmust
beavailableinordertoestablishnormalbehavior.
Figure 6: LEFT‐Position trace of a RO‐RO vessel, color
coded according to the ROT (high ROT‐red, low ROT‐
blue),imposedonamap.Asharpturnaround,withahigh
ROT is visible, which most probably is not an emergency
situation. Taking the time (8th May, 21:53) and the
following short stayin port (∼ 3 minutes) into account,it
rather indicates that the vessel was empty and had a late
customerpickup.RIGHT‐Regulartracksofaferryingray
(overhalfayear),indicatethatthenormalpathis straight
ahead.Anunusualmaneuver,asdetected
byouralgorithm
(colorcodedaccordingtotheROT)isalsoshown.Thismay
indicate that an obstacle (non‐AIS emitting object, e.g.
fishingnet) was encountered,causing correctionof course
comparedtothenormalbehavior.
Table1.Summaryofpotentialnon‐normalnavigationalmaneuverswithhighheeling.Thecollisioneventonthe16th.June
sticks outwith a veryhigh ROT anda high CA. Theisland passage onthe 15th. June, however, musthave felt almost
equallyuncomfortable.Noneoftheothercandidatesarein
thevicinitytotheseevents.
_________________________________________________________________________________
Date |ROT|max |CA|max Maneuver Avg.wave Comment
[deg/min] [m/s
2
]duration[s] height[m]
_________________________________________________________________________________
20thFeb. 155.40.3662.6Yawinginheavysea
20thMar. 150.20.34491.9Turninheavysea
10thApr. 145.00.3742.3Sharpturninheavysea
16thApr. 160.70.36113.4Yawinginheavysea
15thMay 155.40.28712.0Sharpturninheavy
sea
15thJun. 177.20.45601.0Sharpturn
16thJun. 194.50.44600.8Collision
_________________________________________________________________________________
75
3.3 DependencebetweenROTandCAonseastate
Obviously, a small fishing vessel will also be more
exposedtoroughseas,requiringmorerudderactions
thanalargefreighter.ThecharacteristicsoftheROT
values may therefore also show seasonal
dependencies.Theeffectofwavesontheferry’sROT
fluctuations,previouslydescribedinourcasestudy,
is given in Figure 8 below. Here it seems that the
large rudder actions stem from navigating in high
seas. Especially in situations when a vessel
encounters a following or quartering sea, a
phenomenoncalledyawingoccurs,wherethevessel
exhibitssideto
sideturningrequiringlargeROTsto
stayoncourse.Thisisexactlywhatcanbeobserved
intheROTandSOGseries shown in Figure 8. The
centripetal acceleration CA is proportional to the
productbetweenROTandSOG,asseenfromEq.(4).
Thus,inaroughsea,bothROTand
SOGwilldepend
on the sea state. It is therefore interesting to
investigate if the rate of change in CA (time
derivative) gives an indication of the sea state
condition. The time series in Figure 9 shows the
maximum wave height, measured over four daily
time intervals, together with the
maximum of the
time derivative of CA, over the same time period.
OnlyROTandCAsamplesstemmingfromvoyages
were used, i.e. all harbor stays were excluded.
Interestingly, there is indeed a significant co‐
variation(dependency)betweenthemaximumwave
heightandthemaximumrateofchangeinCA.The
smoothed lines in Figure 9, is found to have a
correlation coefficient of ∼ 0.6. This observation
could indicate that the AIS signal may actually be
used to derive information about the currently
occurringseastate.Thisinformationcould,insome
cases, be used to rule out some of the
potential
candidates for critical maneuvering, since some of
the high ROT values could then be ascribed to
maneuveringinheavyseas.
Figure8. There was a rough sea on 16th of April with a
recordedaveragewaveheightof 3.4m.LEFTRate ofturn
plottedversustime.Largerudderactionscompensatingfor
the impact of the waves are visible. A single maneuver
caused an extreme ROT value of magnitude 160.7
[deg/min], as
indicated by the red dashed vertical line.
RIGHT‐The variation in speed over ground with time,
showstheferrys fatiguingcoursethroughthewaves. The
red dashed vertical line corresponds to the time of the
extremeROTvalue,shownintheleftpanel.
Figure9.Normalizedmaximumwaveheight(at4different
timesaday,bluetriangles)andnormalizedmaximumrate
ofchangeinCA(atthesametimeperiodofday,reddots),
together withfittedsmoothing splines in the same colors
(smoothing parameter 0.2). Observe the nicely visible co‐
variation of the
smoothed signals. The gray bars at the
bottomindicatethenumberofAISsamplesavailableinthe
varioustimeperiods(max.9700samples/period).
4 CHALLENGES
4.1 Criticalsituationsdoesnotalwaysimplyextreme
ROT
A fundamental challenge with our methodology, is
the fact that only critical situations where non‐
normal maneuvers have been performed can be
identified.Incaseswerethenavigatorfellasleep,ora
situation was erroneously considered as not critical
or
insuddengroundings, noevasivemaneuverswill
have occurred and therefore no non‐normal ROT
signalswillbeobservable.Toidentifythesecasesone
may have to fall back to the traditional (and
time/resource intensive) approach of analyzing
overlapping zones between ships. In cases of
groundingsasuddendropof
thespeedoverground
to zero are relatively easy to identify. The reader
mustalsorealizethattheROTwillbevesselspecific,
i.e.smallervesselslikefishingboatsaremuchmore
maneuverable and will therefore exhibit normally
largerROTs(andCAs).Ontheotherhand,forlarge
vesselsitmay
bealmostimpossibletoseelargeROTs
duetotheirinertiaandmomentum. This means, in
ordertodeterminewhetheranon‐normalmaneuver
hasbeenperformedoneshouldnotbase thisdecision
onsingleROTvaluesinisolation,one has to know
thenormalbehaviorforthatvessel.Once
normality
isestablished,fromrepresentativehistoricROTdata,
a given ROT can then be benchmarked against the
normalbehaviorforthatspecificvessel(orgroupof
similar vessels). It should also be pointed out that
duringthemaneuverabilitytestsofanewvessel,as
demandedbyIMO,anupper|ROT|value,
validin
calmsea,willberecorded. If,during anoperation,a
ROTshowstobeclosetothisupperlimit,onemay
anticipatethatindeedanextraordinaryrudderaction
has been performed. We also showed that a high
ROTcanoccurinroughseas.Anindicationaboutthe
sea
state can be derived by computing the rate of
changeinCA. Further verification of these findings
andamoreelaboratestudyislefttofuturework.
76
4.2 TheBigDataproblem
InSection2.1wedemonstratedbyacasestudyhow
theROTparameterinahighresolutionAISsignalfor
a single ship can be used to identify non‐normal
behavior. Certainly, this path is always faster
compared to more standard approaches of
identifying near
collisions, i.e. computing the
minimum geographical distance of one vessel to its
nearest neighbors in time. The success of the ROT
approach assumes that an evasive maneuver was
initiatedandthattheROTindicatorisproperlysetup
withtheAIStransponder.
Identifyingvaluesabove athresholdfora single
time
seriesof2millionpointsisstraightforwardona
regular desktop computer. Also for a small ship
owner, it is not very challenging to handle full
resolution AIS data for his fleet. However, our test
dataset contains records for roughly 16,000 vessels
forhalfayear.Ifthisisthe
startingpoint,andmaybe
expandedtocoveratimespanofseveralyears,itis
notstraightforwardtostoreandhandleallthedata
due to its volume. To overcome this challenge, the
AIS signals have been stored in Apache HBase
(scalable no‐SQL database) on a Hadoop cluster,
allowing fast and scalable analysis on shore. The
examplecaseconsideredinthisstudywasidentified
bothtroughlookingatROTvaluesalone,amongstall
vessels,andbyrunningamorecomplexalgorithmto
identify minimum distance between nearest
neighbors.Wedeferfurtherdetails about this work
to a forthcoming publication,
as it is outside the
scopeofthispaper.
4.3 ROTiscurrentlynotareliableparameterintheAIS
feed
Themaindrawbackliesinthefactthatveryoftenthe
rateofturnindicator isnotconnectedwiththeAIS
transponder,althoughaccordingtoSOLASSectionV
Reg
19,(2.9)”Allshipsof50000grosstonnageand
upwardsshall,...have:(2.9.1)a rateofturnindicator,
orothermeans,todetermineanddisplaytherateof
turn; and (2.9.2) a speed and distance measuring
device, or other means, to indicate speed and
distance over the ground. ...
If a ship is equipped
withanAISsystemandarateofturnindicator,then
the ... . Rate of Turn values are hold in the
corresponding AIS data field.” Unfortunately, the
majorityofshipsinthedatabasefromtheNorwegian
Coastal Authority donotlog(ortransmit correctly)
theirrateofturnintheAISsignal.Fromroughly16
000shipsinourdataset,only6.5%oftheshipshad
20(ormore)differentrateofturnvalues,whereasthe
vast majority had less than just 4 different values.
Thisisinalignmentwithqualityassessmentstudies
of
AISsignals(Felski&Jaskolski(2012)).
4.4 ReconstructionofROTfromotherparameters
In case of missing or bad values, the ROT and CA
mayinprinciple be reconstructedfromtheheading
or trace and speed of the vessel, e.g. Aarsæther &
Moan(2007).However,ifthe AISsamplinginterval
is too long, the ROT can no longer reliably
reconstructed from the heading signal.
Reconstruction testsperformedshowedthat,dueto
thelargeinertiaofvessels,anyshorttermlargeROT
values are effectively smoothed out. On the other
hand, the use of bow propeller or tug boats for
maneuvering can
result in sharp turns in the AIS
traces, which are then wrongly interpreted as
instances with high ROT. It is not only the lack of
ROT values in the AIS feed that is challenging, but
also the data quality issues when trying to
reconstruct ROT or CA. Known and experienced
examples are: 1) the AIS longitude and latitude
valuesmayjump severalhundredmeters,2)wrong
timestampsareassigned,and3)acontinuoustrace
willsoonerorlaterhaveholes(lackofdata).
5 CONCLUSION
In this paper we presented a new methodology for
identifying and characterizing occurrences of non
‐
normal maneuvers, that could be candidates for
safety critical situations. The method utilizes high
frequency AIS data feeds, and utilizes the usually
disregarded rate of turn and speed over ground
parameters.AveryhighROTvaluemayindicatean
unnatural large change in the heading of the ship.
Our central
assumption is that most safety critical
situationssuchas(near)collisionsarealmostalways
accompaniedbysomemoreorlesssuccessfulevasive
maneuvers, i.e. maneuvers with sharp turns. This
approach requires the capability of handling and
analyzing large amount of data. For instance, the
entire data set for all vessels in
Norwegian waters
overahalfayear,consistsofroughly3billionAIS
records.
Our study clearly demonstrates that from high
resolution ROT, it is straight forward and indeed
very fast to single out potential critical situations
characterized by non‐normal maneuvers. Knowing
the place and time of the incidence,
it remains to
checkiftherewereanyanothervesselsinthevicinity
by e.g. applying the zone approach. It is also
importanttorealizethatnon‐normalmaneuversmay
occurevenwhentherearenoothervesselsnearby.
The examples in Figure 6 demonstrate that a
significant amount of human
behavior can be
inferred from ROT and SOG data. Note that our
methodology automatically takes into account how
dangerousasituationisperceivedbythenavigator.
Themoredangerous anavigatorseesan encounter,
themorevigoroushewillmaneuver,andthelogged
ROT will consequently be higher. One could
also
studyhowevasiveactions from ROTandSOG will
relatetotheshortestdistancebetweenvessels,orone
may even show what actions were not taken. For
instance, in our case study, it was found that the
evasive maneuver started 31.7 seconds before the
collision. A similar study may be
performed to
examinewhennavigatorsstarttoinitiatetheircourse
adjustments to avoid critical situations. In the end,
individual navigators (within a shipping company)
couldbebenchmarkedregardinghowriskavertthey
arecomparedtoothers(i.e.ROT) and what istheir
comfortprofile(i.e.CAandimpulse).
As AIS
signals are continuously received by the
Vessel Traffic Stations (VTS), an occurring non‐
normal evasive maneuver could be detected in real
77
time and flagged on their ECDIS (Electronic Chart
Display and Information System) displays. This
would of course not avoid critical situations, but it
might force navigators to behave more carefully as
theyknowthattheVTScanseewhenandhowthey
reacted.
The intention of this paper has been
twofold; to
demonstrate the value of so far disregarded
parameters,e.g.ROT,intheAISfeed,andtopresent
a new way of identifying and analyzing potential
safety critical situations using ROT instead of a
traditionalzoneconcept.Ourgoalistoconvincethe
readerthatthereareindeedother
parametersinthe
AIS data feed that could give valuable information,
other than the geo location. We believe the rate of
turnisofspecial interest,asitalmostallowslooking
thenavigatoroverhisshoulder,andseewhathe/she
is doing (or not doing). We hope that our work
initiates
moreresearchonsafetyatsea,thatactually
usesmeasurementdatasuchasROTinAIS.Research
onhumanbehaviorinvessel maneuvering is based
ondifferentbehavioralstrategiessuchasriskprone,
average risk, and neutral risk (Hoogendoorn et al.
(2013)) which could actually be correlated to the
rudder
actions(ROT).
OurmessagetotheIMOandnationalauthorities
is to make sure ROT is logged and transmitted
correctly.Thecurrentfraction of around 5 % of the
ships,properlyloggingROT,istoolow.
ACKNOWLEDGEMENTS
The authors wish to acknowledge the Norwegian
Coastal Authority (Kystverket), and especially
Harald ˚
Asheim for providing us with the high
resolutionAISdata.
REFERENCES
Aarsæther,K.,&Moan,T.(2007).Combinedmaneuvering
analysis, ais and full‐mission simulation. International
Journal on Marine Navigation and Safety of Sea
Transportation,1,31–26.
Bertram,V.(2000).PracticalShipHydrodynamics.Oxford:
Butterworth‐Heinemann.
Coldwell, T. (1983). Marine traffic behaviour in restricted
waters.TheJournalofNavigation,36,
431–444.
Felski,A.,&Jaskolski,K.(2012).Informationunfitnessasa
factorconstrainingautomaticidentificationsystem(ais)
application to anti‐collision manoeuvring. POLISH
MARITIMERESEARCH,19,60–64.
Fujii, Y., & Tanaka, K. (1971). Traffic capacity. Journal of
Navigation, 24, 543–552. URL:
http://journals.cambridge.org/article
S0373463300022384.doi:10.1017/S0373463300022384.
Goerlandt,F.,Montewka,J.,
SonneRavn,E.,Hanninen,M.,
&Mazaheri,A.(2012).EfficienSea‐analysisofthenear‐
collisions using AIS data for the selected locations in
theBalticSea.URL:http://www.efficiensea.org.
Goodwin,E.M.(1975).Astatisticalstudyofshipdomains.
Journal of Navigation, 28 , 328–344. URL:
http://journals.cambridge.org/article
S0373463300041230.doi:10.1017/S0373463300041230.
Hoogendoorn,S.,Daamen,
W.,&Shu,Y.(2013).Modeling
human behavior in vessel maneuver simulation by
optimal control and game theory. Transportation
Research Record: Journal of the Transportation
ResearchBoard,2326,45–53.
Kao, S., Lee, K.‐T., Chang, K.‐Y., & Ko, M.‐D. (2007). A
fuzzy logic method for collision
avoidance in vessel
trafficservice.TheJournalofNavigation,60,17–31.
Llana, C.B., &Wisneskey, G.P. (1991). Handbookof the
nauticalrulesoftheroad.NavalInstitutePress.
Mestl, T., Anderssen, T., Zachrisen, M., & Hjøllo, B. ˚A.
(2008).Towardsthecollaborativesea.EuropeanJournal
of
Navigation,6,14–16.
Montewka,J.,Goerlandt,F.,Lammi,H.,&Kujala,P.(2011).
A method for assaessing a causation factor for a
geometrical mdtc model for ship‐ship collision
probabilityestimation.TransNav,5,365–373.
Pedersen, P. (1995). Collision and grounding mechanics.
The Danish Society of Naval Architects and
Marine
Engineers,(pp.125–157).
Pietrzykowski,Z.(2008).Shipsfuzzydomainacriterionfor
navigationalsafetyin narrowfairways.The Journalof
Navigation,62,499–514.
Pietrzykowski, Z., & Uriasz, J. (2009). The ship domaina
criterion of navigational safety assessment in an open
seaarea.TheJournalofNavigation,61
,93–108.
Rawson, K. J., & Tupper, E. C. (2001). Basic ship theory:
shipdynamicsanddesignVOL.2.Oxford:Butterworth
‐Heinemann.
Sturt,R.H.B.(1991).Thecollisionregulations.NewYork,
N.Y:Lloyd’sofLondonPress.
IMO (1972). COLREG preventing collisions at sea. URL:
http://www.imo.org/en/OurWork/Safety/Navigation/Pa
ges/Preventing‐Collisions.aspx.
IMO
(2003).Guidelines forthe installationof a shipborne
automatic identification system (AIS). URL:
http://www.imo.org/en/OurWork/Safety/Navigation/D
ocuments/227.pdf.
IMO(2008).MSC‐MEPC.7/Circ.7‐GuidanceOnNear‐Miss
Reporting. URL:
http://www.safety4sea.com/images/media/pdf/2010.10.2
5 IMO%20Guidance%20‐%20near‐misses%20‐%20MSC‐
MEPC.7.pdf.
IMO (2015). IMO Sub‐committee on navigation,
Communications and Search and Rescue‐a study on
ship operator centred collision prevention
and alarm
system.URL:NCSR2/INF.10.
The Transportation Safety Board of Canada (1998). Near‐
collision,betweenthecruiseship”STATENDAM”and
the tug/barge unit ”BELLEISLE SOUND”/”RADIUM
622”, Marine Occurrence Report M96W0187. URL:
http://www.bst‐tsb.gc.ca/eng/rapports‐
reports/marine/1996/m96w0187/m96w0187.pdf.
Wang,N.,Meng,X.,Xu,Q.,&Wang,Z.(2009). Aunified
analyticalframeworkforship
domains.TheJournalof
Navigation,62,643–655.
Wikipedia (2015). Automatic Identification System. URL:
https://en.wikipedia.org/wiki/Automatic Identification
System.
Zhang,W.,Goerlandt,F.,Montewka,J.,&Kujala,P.(2015).
A method for detecting possible near miss ship
collisionsfromAISdata.Ocean Engineering,107,60–
69.
Zhao, J., Wu, Z., & Wang,
F. (1993). Comments on ship
domains.TheJournalofNavigation,46,422–436.
