1 / 14
文档名称:

Capgemini-海洋数据和人工智能用于物种保护(英)-2022.10-14正式版.doc

格式:doc   大小:12,489KB   页数:14页
下载后只包含 1 个 DOC 格式的文档,没有任何的图纸或源代码,查看文件列表

如果您已付费下载过本站文档,您可以点这里二次下载

分享

预览

Capgemini-海洋数据和人工智能用于物种保护(英)-2022.10-14正式版.doc

上传人:youyicheng 2022/11/24 文件大小:12.20 MB

下载得到文件列表

Capgemini-海洋数据和人工智能用于物种保护(英)-2022.10-14正式版.doc

文档介绍

文档介绍:该【Capgemini-海洋数据和人工智能用于物种保护(英)-2022.10-14正式版 】是由【youyicheng】上传分享,文档一共【14】页,该文档可以免费在线阅读,需要了解更多关于【Capgemini-海洋数据和人工智能用于物种保护(英)-2022.10-14正式版 】的内容,可以使用淘豆网的站内搜索功能,选择自己适合的文档,以下文字是截取该文章内的部分文字,如需要获得完整电子版,请下载此文档到您的设备,方便您编辑和打印。OceanData
andAIfor
species
conservation
1
Abstract.
Theproblemandatthesametimeour
motivationisthelossofspeciesdiversityinthe
ocean,whichisoftenalsoreferredtoas”invisible
dying”.Withourapproach,wepursuethegoalof
makingthisdyingvisibleandthuspreventable.
Tomakeeventsintheoceanvisible,weneed
toidentifypatternsdependingontheocean
depthandthenrecognizedeviationsfromthese
-
Vesteralen,weareanalyzingoceandata,tohelp


shouldhelptoidentifytheconsequencesof
humaninterventioninnature,suchasdwindling
fishstocks.
OCEANDATAANDAIFORSPECIESCONSERVATIONIOCTOBER2022 2
Introduction
Accurateobservationofecosystemsenabledetailedoceanographicresearch,allowinganomaliestobeidentifiedinenormousamountsofdatawiththehelpofartificialintelligence(AI).
TheLofoten-Vester˚alen(LoVe)OceanObservatoryislocatedwestofHovdenVester˚,geological,oceanographicandeconomic”hotspot”.Anetworkofsubmarinecablesandsevensensornodescoversacross--basedstationandsevensensorplatforms,,andhasbeenactivesince2013.
Thesystemisboth,anationalresearchinfrastructure,basicandappliedresearch,aswellasatestinfrastructure,-Vester˚alenOceanObservatoryhascollectedover100terabytesofsensordata(temperatures,currents,echograms)overtheyears.
Theteam
ThomasRamm
.
MajedAlaitwni
.
GeirPedersen
isaresearcherattheLoVeOceanOceanObservatoryandsupportstheprojectontheNorwegianside.
TomHatton
isaDataScientist,hismaster’sthesisexploredtheuseofunsupervisedAImodelsforanomalydetectioninhigh-.

SophieBader
isamolecularbiologistspecializinginoceanicecosystems,.
parMustaphaMustapha
.
DanielFriedmann
.
NilsOlavHandegaard
.
,.
OCEANDATAANDAIFORSPECIESCONSERVATIONIOCTOBER2022 3
OceAInwascreatedasateamnametoparticipateinCapgemini’sGlobalDataSciencechallenge(GDSC).
(1)directionalpulsatingsoundsinspecificareaswithascientificechosounder,(2)aso-calledhydrophone,,(3)anAcousticDoppler(ADCP)thatdetectsthe
,witheachtypeofdata(hydrophone,biomassdetection,etc.)
ApacheAirflow,whichworksonthe”ConfigurationAsCode”(DirectedAcyclicGraph).ItisalsoworthmentioningthatDockercontainersareactuallymanagedbyKubernetes,,,whicharedisplayedwiththedataintheformofimagefilesthroughaninteractivewebinterface.

,,sometypesofanomaliesareonlydetectableifthedatapointsareconsideredinterconnectedfromthebeginning.
,thisapproachshowedsomedisadvantages,suchasincreasedcomplexityinaggregatingtheindividualmodels.
speedanddirectionofoceancurrentsusingtheDopplereffect,andfinally,(4)pointsensorsthatprovidereal-timephysical,biologicalandchemicalobservations.
,suchasdifferencesinfishpopulations,varyingcurrentpatterns,ortheinfluenceofclimatechange.
Whilelargevolumesofrawdataaredifficulttoprocessmanuallyandtheresultsarehighlyerror-proneintheprocess,,AIenablescontinuousanalyzationofincomingdata,resultinginastreamofdatatotheresearchers.
ArchitekturdesSystemsEventhoughtheAImodelisthecoreofOceAIn,-nativearchitectureconcept,whichwillcreateafuture-


MODEL
F1SCORE
PRECISION
RECALL
MACRO
AVGF1
alwaystrue




alwaysfalse




uniformrandom




stratified




OCEANDATAANDAIFORSPECIESCONSERVATIONIOCTOBER2022 4
Theideathatultimatelyprevailedwastouseadeep-,unsupervisedmodelsweresupposedtobeused,butthisapproachturnedoutnottobefeasible.
TheuseofunsupervisedAImodelshastheadvantagethatthetrainingofthosemodelsdoesnotrelyonthepresenceoflabeledtrainingdata().
However,thosemodelsareextremelyvulnerabletodatanoiseandcorruption[1].Theunderlyingoceanmeasurementdataarealsosubjecttothesecharacteristics,whicharefurtherexacerbatedbytheirwhole-scalenatureandthehighdatadimensionalitythataccompaniesthem.
,supervisedAImodelsoutperformtheirunsupervisedcounterpartsinanomalydetection,astheyareparticularlycapableofdetectingapplication-specificanomalies[2].
Tochecktheperformanceoftheunsupervisedmodelslater,fourbaselinemodelswerecreatedfirst,

one-dimensionalconvolutionsineachofthehiddenlayers,-calledLSTMlayersandthethirdisafullyconnectedautoencoder.

RECONSTRUCTION
CLEANDATA
FULLDATA
dense
lstm
conv
lstm
conv
denseW32
lstmW32HL100
denseW16
lstmW32HL100
convW16
HL40040
40204
HL10040204
40204
HL100010010
denseW32
lstmW32HL100
denseW32
lstmW32HL100
convW16
HL40040v2
40204v2
HL10040204
40204v2
HL1000400100
4
denseW8HL400
denseW32
lstmW32HL40
convW16
40204
HL20001000
lstmW32HL400
HL10040204
10040204
10010
convW32
denseW32
lstmW32HL64
HL10040204
HL40040
3216
convW64
denseW32
lstmW8HL100
HL10040204
HL40040v3
40204
ofmachinelearningmodelsandarenaivesolutionsforaclassificationproblem.
Bycomparingtheresultsofarealmodelwiththoseofthebaselinemodels,,onealwaysclassifiesfalse,onealwaysclassifiestrue,onesplitsthedatasetsinhalfbetweentrueandfalse,andthelaststratifiesthedatabasedontheirlabels.
Duringthemodeldevelopment,,

,aswellasanarchitecturethatusesconvolutionandmax-,”W”andanumberisgiven,whichrepresentsthewindowsize.”W32”,,thereisalsoaversionnumberattheend,whichindicatesmodelsthathaddeliveredpromisingresultsinthefirstinstance,,thereisanother

,,therewasonlyonemodel,”lstmW32HL128
128128”whichcould(minimally),while4showstheaverageresultsofthedifferenttypesofmodels.
,noise,anddatacorruptiongreatlydegradetheresultsofunsupervisedmodels.
OCEANDATAANDAIFORSPECIESCONSERVATIONIOCTOBER2022 5

RECONSTRUCTION
CLEANDATA
FULLDATA
dense
lstm
conv
lstm
conv
denseW32
lstmW32HL128
convW32
lstmW32HL128
convW32HL32
HL256*5v2
128128
HL128128128
128128
3232
convW32
HL128128128v2
convW32HL32
3232

MODEL
QUANTILE
F1SCORE
PRECISION
RECALL
MACROAVGF1
reconstructiondensecleandata





reconstructionlstmcleandat





reconstructionconvfulldata





reconstructiondensefulldata





reconstructionlstmfulldata





forecastconvcleandata





forecastdensecleandata





forecastlstmcleandata





forecastlstmfulldata





forecastconvfulldata





OCEANDATAANDAIFORSPECIESCONSERVATIONIOCTOBER2022 6
Inaddition,thedatasetsarehighlydimensional,,,,,.

Forallthesereasons,,.
,,whilealldatapointsthataredetectedasdeviationsarestoredinthearrayanomalous˙data˙indices.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

=test_mae_loss>threshold
print("Numberofanomalysamples:",(anomalies))
print("Indicesofanomalysamples:",(anoma