建筑结构的生成式智能设计方法研究进展 Generative AI design for building structures_第1页
建筑结构的生成式智能设计方法研究进展 Generative AI design for building structures_第2页
建筑结构的生成式智能设计方法研究进展 Generative AI design for building structures_第3页
建筑结构的生成式智能设计方法研究进展 Generative AI design for building structures_第4页
建筑结构的生成式智能设计方法研究进展 Generative AI design for building structures_第5页
已阅读5页,还剩33页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

checkfor

updates

AutomationinConstruction157(2024)105187

Contentslistsavailableat

ScienceDirect

AutomationinConstruction

journalhomepage:

/locate/autcon

Review

GenerativeAIdesignforbuildingstructures

WenjieLiao

aXinzhengLua,,

*,YifanFei

b,YiGub,YuliHuang

a

aKeyLaboratoryofCivilEngineeringSafetyandDurabilityofMinistryofEducation,TsinghuaUniversity,Beijing100084,China

bBeijingEngineeringResearchCenterofSteelandConcreteCompositeStructures,TsinghuaUniversity,Beijing100084,China

ARTICLEINFO

ABSTRACT

Keywords:

Designingbuildingstructurespresentsvariouschallenges,includinginefficientdesignprocesses,limiteddata

Buildingstructuraldesign

reuse,andtheunderutilizationofpreviousdesignexperience.Generativeartificialintelligence(AI)hasemerged

Datafeaturerepresentation

GenerativeAIalgorithm

Designevaluation

Intelligentoptimization

asapowerfultoolforlearningandcreativelyusingexistingdatatogeneratenewdesignideas.Learningfrompastexperiences,thistechniquecananalyzecomplexstructuraldrawings,combinerequirementtexts,integratemechanicalandempiricalknowledge,andcreatefreshdesigns.Inthispaper,acomprehensivereviewofrecentresearchandapplicationsofgenerativeAIinbuildingstructuraldesignisprovided.Thefocusisonhowdataisrepresented,howintelligentgenerationalgorithmsareconstructed,methodsforevaluatingdesigns,andtheintegrationofgenerationandoptimization.ThisreviewrevealsthesignificantprogressgenerativeAIhasmadeinbuildingstructuraldesign,whilealsohighlightingthekeychallengesandprospects.Thegoalistoprovideareferencethatcanhelpguidethetransitiontowardsmoreintelligentdesignprocesses.

1.Introduction

Thedesignofbuildingstructuresisanuancedtaskthatnecessitatestheblendingofempiricalandmechanicalknowledge.Engineershavebeenconsistentlydiscovering,developing,andimplementingsophisti-catedcomputer-aideddesigntechnologiestostreamlinetheprocessandtargetefficientandreliablestructuraldesigns.

Inthe1980s,intelligentdesignmethodswereproposedbasedonexpertsystemalgorithms[

62

].Inthefollowingyears,aseriesofintel-ligentdesignmethodsbasedonbiologicallyinspiredalgorithmsemergedtogetherwiththeconceptofgenerativedesign[

80

,97

,

102

].Advancementsincomputertechnologydrovethedigitizationandautomationofbuildingstructuraldesignsforwardatanunprecedentedpace.However,expertsystemalgorithmsandbiologicallyinspiredal-gorithmsfindthemselvesgrapplingwithissuesindatalearning,designruleencoding,anddesignefficiency,whichhampertheirwiderapplication.

Inrecentyears,artificialintelligence(AI)technologies,notablydeeplearning,havemaderemarkableprogressinlearningfromexistingdataandgeneratingnewdesigns,whichsetsthemapartfromconventionalexpertsystemsandbiologicallyinspiredalgorithms.GenerativeAIdesign,whichharnessesmachinelearningalgorithmstolearnfromdataandunfoldsnewcontent,hasbeenidentifiedasoneofthetoptentechnologytrendsfor2023[

12

].Intelligentgenerativetechnologies,

suchasDALL-E[

8

]andChatGPTbyOpenAIandAlphaFoldbyDeep-Mind[

41

],havedemonstratedtheversatilityofgenerativeAIinvariousfieldsandhavebecomeacutting-edgeareaofresearch.

Buildingstructuraldesignsrelyprimarilyondrawingsthattranslateintostructuredorimagedata.Engineerscompletethedesignprocessbasedonanarchitecturaldesignwithmultipleconstraintssuchascomplianceandeconomy.Theend-to-enddesignprocessofgenerativeAIisconsistentwiththatofbuildingstructuraldesignbyengineers[

104

],equippedwithpowerfullearningandgeneratingcapabilitiestotackletheintricatepuzzlesofintelligentdesign.Therefore,ithasbecomeanewresearchtopic.

SeveralcomprehensivereviewshavebeenundertakentoexploretherecentadvancementsinintelligenttechnologiesbasedonAIinbuildingstructuraldesign,analysis,construction,andmaintenance(

Table1

).ThesereviewscontributesignificantlytotheunderstandingofvariousAItechnologies,includingmachinelearning,deeplearning,GenerativeAdversarialNetworks(GANs),GenerativePre-trainedTransformer(GPT)models,andGraphNeuralNetworks(GNNs).Bydelvingintothesetopics,thesereviewsshedlightonthepotentialapplicationsandbenefitsoftheseAItechnologiesinbuildingstructures.

Intherealmofdesignandanalysis,severalnotablereviewstudieshavebeenconducted.AldwaikandAdeli

[59

]conductedareviewfocusingontheoptimizationofhigh-risebuildingstructures,utilizingnature-basedoptimizationapproachesliketheneuraldynamicsmodel

*Correspondingauthor.

E-mailaddress:

luxz@

(X.Lu).

/10.1016/j.autcon.2023.105187

Received11June2023;Receivedinrevisedform15October2023;Accepted5November2023

Availableonline11November2023

0926-5805/©2023ElsevierB.V.Allrightsreserved.

2

W.Liaoetal.

Table1

ScopesofexistingreviewstudiesrelatedtoAI-assistantbuildingstructuraldesign.

Dsgn.

Anlys.

C&O&M

DL

(inc.

Gen.

AI)

ML

Optim.

DRL

Aldwaik&

Adeli[

59

]Chietal.

[37

]Amezquita-

Sanchez

etal.[

49

]

LiandAdeli

[

120

]

Afzaletal.

[

58

]

Sunetal.

[35

]Pizarroetal.

[

77

]

Badugeetal.

[

95

]

Wuetal.[

11

]Zakian&

Kaveh[

69

]Yükseletal.

[

65

]

Wangetal.

[

19

]

Omranyetal.

[

34

]

Sakaetal.

[9

]Jiaetal.

[110

]Topuz&

ÇakiciAlp

[

14

]

Koetal.[

42

]Ours

Dsgn.=Design,Anlys.=Analysis,C&O&M=Construction&Operations&Maintenance,DL=DeepLearning,Gen.AI=GenerativeAI,ML=MachineLearning,Optim.=Optimization,andDRL=DeepReinforcementLearning.

andgeneticalgorithms.Chietal.[

37

]andOmranyetal.[

34

]exploredtheintegrationofbuildinginformationmodeling(BIM)withsmarttechnologiesinstructuraldesignpractice,suchasstructuraldesign,planning,analysis,andoptimization.Afzaletal.[

58

]providedanoverviewofstructuralcomponents,optimizationstrategies,andtheutilizationofvariouscomputationaltoolsinRCstructuraldesignopti-mization.Pizarroetal.

[77

]reviewedrule-andlearning-basedmethodsforintelligentrecognitionanddesigninarchitecture.Sunetal.

[35

]delvedintotheresearchonmachinelearningapplicationsinbuildingstructuredesignandperformanceevaluation.ZakianandKaveh[

69

]presentedanoverviewofseismicdesignoptimization,encompassingcommonsolutionmethods,optimizationproblemtypes,andoptimiza-tiongoals.Yükseletal.

[65

]analyzedtheresearchstatusofgenetical-gorithms,fuzzylogic,andmachinelearningmethodsinengineeringstructuredesign,coveringareassuchasdesigngeneration,evaluation,optimization,decision-making,andmodeling.Wangetal.[

19

]con-ductedareviewspecificallyontheapplicationofAItechnologyinmaterialandstructuralanalyseswithinthefieldofcivilengineering.TopuzandÇakiciAlp

[14

]evaluatedthecurrentstateofmachinelearningincomputer-aideddesign,engineering,andmanufacturingofarchitecture.Lastly,Koetal.

[42

]introducedadvancementsinauto-matedspatiallayoutplanningcombinedwithAI.

SeveralreviewstudiesalsoexploredAIapplicationsintherealmofconstructionandmaintenance.Amezquita-Sanchezetal.[

49

],andLiandAdeli[

120

]providedaninsightfulreviewandoutlookontheuti-lizationofmachinelearningtechnologiesinstructuralsystemidentifi-cation,healthmonitoring,vibrationcontrol,design,andoptimization.Badugeetal.[

95

]introducedthelatestadvancementsinapplyingAItechnologies,includingmachinelearninganddeeplearning,inbuilding

AutomationinConstruction157(2024)105187

designandvisualizationwithinthecontextofthebuildingindustry4.0.Wuetal.[

11

]presentedacomprehensivestate-of-the-artreviewontheutilizationofGANstotacklechallengingtasksinthebuiltenvironment.Sakaetal.[

9

]assessedthepotentialofGPTmodelsintheconstructionindustry,identifyingopportunitiesfortheirimplementationthroughouttheprojectlifecycle.Jiaetal.

[110

]exploreddiverseapproachesforconstructinggraphdatafromcommonconstructiondatatypesandhighlightedthesignificantpotentialofGNNsfortheconstructionindustry.

However,thereisanoticeablegapintheanalysisanddiscussionsurroundingthedevelopmentandapplicationofgenerativeAIinthestructuraldesignofbuildings.Thisgapresultsinalackofsystematicandtargetedresearchwithinthisfield.Uponreviewingexistingstudies,itbecomesevidentthattheresearchfindingsinthisareaarefragmentedandincomplete.Researchersencounterchallengesinaddressingspecificissues,whiletheentrybarriersfornewresearchersenteringthisfieldareontherise.ThesefactorscollectivelyimpedefurtherprogressandhindertheadvancementofgenerativeAIinbuildingstructuraldesign.Therefore,thisstudyaimstoprovideanoverviewofexistingintelligentstructuraldesignmethodsandfocusesontheachievementsandappli-cationsofgenerativeAIinbuildingstructuraldesign,whichhasun-dergonerapiddevelopmentoverthepastthreeyears.TheanalysisofexistingstudiesrevealsthattheprimaryresearchareasingenerativeAI-basedstructuraldesignencompassdatarepresentation,generational-gorithms,evaluationmethods,andtheintegrationofAIgenerationwithdesignoptimization.Byconductinginvestigationsinthesepertinentdomains,theadvancementofAI-basedbuildingstructuraldesigntech-nologycanbeeffectivelyfacilitated.Accordingly,thisstudywillpre-dominantlycenteraroundshowcasingtheaccomplishmentsofexistingresearchwithinthesespecificareas.

Section2

providesamacro-introductiontocurrentAImethodsandtheirapplicationsinbuildingstructuraldesign.

Sections3–5

reviewthedatafeaturerepresentation,generativealgorithmconstruction,andevaluationmethodsingenerativeAI.

Section6

introducesmethodscouplingintelligentgenerationandoptimization,andSection7presentstypicalengineeringapplicationcasesofintelligentbuildingstructuredesign.

Section8

providesrelevantconclusionsandprospectsforthefuturedevelopmentofthisresearchfield.

Consequently,thisstudyhighlightsthepotentialofgenerativeAIinbuildingstructuraldesignandtheneedforfurtherresearchinthisfield.ThedevelopmentandapplicationofgenerativeAIinstructuraldesigncansignificantlyenhancetheefficiencyandaccuracyofthedesignprocess,leadingtomoresustainableandsaferbuildingstructures.

2.Introductiontoartificialintelligence(AI)methodsandapplicationsinbuildingstructuraldesign

2.1.AImethods

ResearchonAImethodshasbeenconductedforaconsiderableperiod.Inthisstudy,thetimelinewaspartitionedusing2012asareferencepoint,whichmarksthe“deeplearningera”owingtothehighlyeffectiveimageclassificationalgorithmAlexNet[

4

].AccordingtoYükseletal.[

65

],asillustratedin

Fig.1

,mostAImethodsbefore2012werecategorizedasclassicalAImethods,whereasthoseafter2012werereferredtoasmodernAImethods.Thisreviewfocusesprimarilyondeep-learningalgorithmsdevelopedafter2012.

(1)ClassicalAImethods

Representativemethodsincludeknowledge-basedsystemsandbio-logicallyinspiredalgorithms.Knowledge-basedsystemsincludeexpertsystems[

23

],fuzzylogic[

48

],andgenerativedesigngrammars[

1

].Biologicallyinspiredalgorithmsincludegeneticalgorithms[

31

],parti-cleswarmoptimization[

84

],andcellularautomata[

25

].

3

100%

0%

W.Liaoetal.

AutomationinConstruction157(2024)105187

Artificialintelligence(Al)

classicalAl

earnngM)

DeepeargD

Artifcialneura

convolutional

neura

networks

(CNN)

cellularautomata(CA)

Differentialevolution

algorithm(DEA)

particleswarm

ptimization(pso)FuzzYlogic

Recurrent

neura

networks

(RNN/LSTM)

BiologicaIIY-inspired

algorithms

Graph

neura

networks

(GNN)

knowledge-based

systems

Geneticalgorithm

network(ANN)

Expertsystem

ModernA

Machinel

GA

Fig.1.Artificialintelligence(AI)methodswidelyadoptedinbuildingstructuraldesign.

(2)ModernAImethods

advancementshaveprovidedengineerswithenhancedinitialdesignsolutions,elevatingbothdesignefficiencyandquality.Theinsightsfrom

Since2012,theAIfieldhasundergoneacomprehensivedevelop-mentdrivenbydeeplearning,whichcanhandlebigdata,extracthigh-dimensionalfeatures,andsignificantlyimprovelearningfromdata[

111

].Typicaldeeplearningmethodsincludedeepconvolutionalneuralnetworks(CNNs)[

4

,

111

],deepgraphneuralnetworks(GNNs)[

99

],anddeeprecurrentneuralnetwork/longshort-termmemory(RNN/LSTM)[

40

,

90

].Inaddition,advanceddeeplearningalgorithmshavebeendevelopedusingthesemethods,suchasthevariationalautoencoder(VAE)

[24

],generativeadversarialnetwork(GAN)[

38

],transformer[

10

],anddiffusionmodels

[28

].Inrecentyears,generativeAI[

10

,

28

,

38

,

96

]hasbecomearepresentativetechnologicaladvancementindeeplearningresearch.TheemergenceofgenerativeAIordeepgenerativemethodshasopenednewpossibilitiesforintelligentdesign. Presently,modernAImethodshavefoundextensiveapplicationsinvariousengineeringfields,includingarchitecture,mechanicalengi-neering,andaerospaceengineering

[65

].Theyhaveproventobeeffectiveingeneratingarchitecturallayouts,renderingarchitecturalimages[

27

,39

,

56

,57

,63

,

64

,79

,89

,

101,107

],generatingwheelstruc-tures[

92

],andaerodynamicshapesofaircraft[

22

].These

theserelatedstudiesserveascrucialreferencesforthedevelopmentofgenerativeAI-basedintelligentdesigninbuildingstructures.

2.2.ApplicationofAIinbuildingstructuraldesign

Thebuildingstructuraldesignprocesscanbedividedintothreeprimarystages:conceptualdesign(schemedesign),detaileddesign(preliminaryoroptimizationdesign),andconstructiondrawingdesign.Amongthesestages,theconceptualdesignstagesignificantlyimpactsthefinaldesignoutcomeandreliesheavilyondesignexperienceandknowledge(asshownin

Fig.2

).Therefore,intheconceptualdesignstage,AIplaysasignificantroleinthestructuraldesignofbuildings.

Thecoretaskofaconceptualbuildingstructuraldesignisdesigngeneration(alsoknownassynthesis).AccordingtoMaher[

62

],buildingstructuraldesignsynthesismethodsmainlyrefertothegenerationofcorrespondingdesignsguidedbyspecificknowledgefragments,suchasheuristicrulesanddescriptiveframeworks.Inaddition,withthedevelopmentandapplicationofgenerativeAItechnologiesthatcanlearndatafeaturesandgeneratenewdesigns,currentstructuraldesign

Highinfluence

Lowinfluence

Lowexpenditure

Highexpenditure

Result

conceptual

Design

precttime

construction

drawing

design

Fig.2.Influenceofthebuildingstructuraldesignphaseonprojectcost[

15

].

pg(pxg)(x)

GeGneaetreaedaataSaSmapmlepdleaata

pdatpad(atxa)(x)

4

W.Liaoetal.

synthesismethodsmainlyincludeheuristicsearch-based,descriptivegrowth-based,andgenerativeAIlearning-baseddesigns(

Table2

).

Heuristicsearch-baseddesignsprimarilyusebiologicallyinspiredcomputationtechniques,suchasgeneticalgorithms,particleswarmoptimization,andcellularautomata[

2

,

5-7

,

33

,

43-45

,

54

,

61

,80

,

88

,

91

,

93

,

112

,

119

].Descriptivegrowth-baseddesignsprimarilyusegenerativedesigngrammar[

1

,

16

,

26

,53

,

87

,

108

].Theseintelligentdesignmethodshavebeenwidelyappliedinmultipleareasofbuildingstructuraldesign,suchasmulti-schemesearch,comparativeselection,andmaterialoptimization,andhavepromotedtheprogressofdigitali-zationandautomationinstructuraldesign.However,theseclassicalmethodsfacedifficultiesregardingdatalearninganddesignefficiency.Incontrast,generativelearningisanintelligentdesignmethoddomi-natedbygenerativeAI,whichhaspowerfuldesigndatalearningandefficientnewdesigngenerationcapabilitiesandhasbeencontinuouslydevelopingandadvancinginrecentyears.

ThemainarchitectureofthegenerativeAIalgorithmsisshownin

Fig.3

,whichincludesadataset,aneuralnetworkmodel,andalossfunctionmodule.ThedevelopmentandapplicationofgenerativeAIinbuildingstructuraldesignwillmainlyfocusonthesethreeparts,andadetailedanalysisandsummaryarepresentedin

Sections3-5

.

Tofurtherunderstandthecurrentstateofresearchonintelligentstructuraldesign,weconductedasearchoftheWebofScienceCoreCollectionforrelevantpapers.First,asearchwasconductedonAI-basedstructuraldesignmethods,yielding188searchresults.Subsequently,searcheswereconductedonclassicalAI-basedandmodernAI-basedstructuraldesignmethods,resultingin130and74searchresults,respectively.Thesearchformulaearelistedin

Table3

.

Fig.4

(a)showsthechangeinthenumberofresearchpapersonAI-basedbuildingstructuraldesigns,indicatingacontinuousincreaseinintelligentdesignresearchovertheyears,withasignificantincreasesince2020.BycomparingtheclassicalandmodernAImethods,asshownin

Fig.4

(b),itcanbeobservedthatclassicalAIwasthemain-streamresearchmethodforintelligentdesignbefore2012.Since2012,modernAIhasrapidlydeveloped,andtherelatedresearchpapershaveincreasedthreefold;however,somedifferencesstillexistcomparedtoclassicalAImethods.Since2020,generativeAItechnologyhassignifi-cantlyimproveditscapabilitiesandhasbeendevelopingin-depthinthefieldofstructuraldesign,withresearchpapersonparwithclassicalAI

AutomationinConstruction157(2024)105187

methodsandaccountingfor48%ofthetotalresearch,showingatrendofsurpassingclassicalAImethods.

Insummary,byanalyzingtherecenttrendsinintelligentbuildingstructuraldesign,AI-basedmethodsusingdeeplearninghavegraduallybecometheprimaryresearchfocus.Thesemethodsarereferredtobydifferentnames,includingintelligentgenerativedesign,deepgenerativedesign,intelligentdesign,anddeep-learning-baseddesign.However,thecoretechnologyofthesemethodsprimarilyinvolveslearningfromexistingdesigndata,empiricalknowledge,andphysicalprinciplestomastertheabilitytogeneratenewdesignsintelligently.AsthisisconsistentwiththeessenceofgenerativeAI,thisstudycollectivelyreferstothesemethodsasgenerativeAI-basedintelligentdesignsforbuildingstructures.

3.Datafeaturerepresentationanddatasetconstruction

Therepresentationofthedatafeaturesiscrucialinbuildingstruc-turalintelligentdesign.Thisrepresentationisrelatedtothegeometricandmechanicalcharacteristicsofthestructuraldesignandissubjecttothelimitationsofdeep-learningalgorithms.Inbuildingstructuraldesign,datafeaturesmainlyincludetopological,pattern,andsizefea-tures.Topologicalfeaturesrefertothelayoutofstructuralcomponentsinspaceandtheconnectionrelationshipbetweencomponentstoeffectivelyresistoveralllateralandverticalloads.Patternfeaturesrefertothelocalgeometricconfigurationofcomponentsthatmustbeinspecialpatterns(suchasL-shaped,T-shaped,andC-shaped)toresisttheloadofthelocalcomponent.Sizefeaturesrefertothecross-sectionalsizesofstructuralcomponents.

GenerativeAIalgorithmsutilizeconvolutionalneuralnetworks(CNNs)toprocessdataastensors(including1-dimensionalvectors,3-dimensionalimages,andn-dimensionalmatrices).Incontrast,graphneuralnetworks(GNNs)aresuitableforprocessingdataingraphs(nodeandedgerepresentations,with1-dimensionalvectorfeaturesembeddedinnodesandedges).Tensordatacanbetterexpresspatternsandsizefeaturesinthedesign,whereasgraphsaremoreconducivetoexpressingtopologicalfeatures.

Therefore,generativeAIalgorithmscaneffectivelyextractfeaturesandlearnbyappropriatelyrepresentingthedatafeaturesinbuildingstructuraldesign.Thissectionreviewsthedatafeaturerepresentation

Table2

Comparisonofthreemainstreamintelligentdesignmodes.

Heuristicsearch-basedDescriptivegrowth-basedGenerativeAIlearning-based

Concept

description

Concept

illustration

Utilizingtheprovidedinitialsolutionwithinpredeterminedboundaryconstraints,the

iterativeprocessaimstosystematicallysearchfortheoptimalresult.

Theprocessinvolvesestablishinggenerationrulesandconstraints,commencingfromtheinitial

design,andsubsequentlyiterativelyoptimizingitbasedonthedescriptivegenerationrules.

Byleveragingexistingdata,AIcomprehendsthe

mappingpatternsbetweenbuildingandstructure.

TrainedAIgeneratesacorrespondingstructuraldesignforanewbuildingdesigninasinglecomprehensive

step.

Representative

methods

Geneticalgorithms

Generativedesigngrammar

GAN

Initialstructuraldesign

Required

Partialrequired

Notrequired

Manually

definedrules

Required

Required

Partialrequired

Iterationdesign

Required

Required

Notrequired

Learning

Notrequired

Notrequired

Required

Performance

High

Relativelyhigh

Medium

Efficiency

Relativelylow

Medium

High

5

Input

output

physical

constraint

Dataloss

W.Liaoetal.

AutomationinConstruction157(2024)105187

Datasets

Neuralnetwork

Evaluation

Groundtruthdata

embeddedphysicallaws

andempiricalrules

Discriminator

Rea

Fake

Generativeneuralnetwork

Fig.3.AlgorithmframeworkofgenerativeAI.

Table3

SearchobjectsandformulasusedintheWebofSciencecorecollectionandtheresultingnumber(accessedonApril20,2023).

Searchobjects

Searchformulas

Numberof

studies

ClassicalAI-based

buildingstructuraldesign

(TS=(“building”OR“architecture*”)

ANDTS=(“structur*design”)ANDTS=

(“intelligen*”OR“automat*”OR

“artificialintelligence”OR“design

intelligence”OR“generative”OR

“optimiz*”OR“explorat*”)ANDTS= (“expertsystem*”OR“fuzzylogic”OR“geneticalgorithm”OR“generative

grammar”OR“evolution”OR“particleswarmoptimization”OR“cellular

automata”))

(TS=(“building”OR“architecture*”)

ANDTS=(“structur*design”)ANDTS=(“intelligen*”OR“automat*”OR

“artificialintelligence”OR“design

130

ModernAI-based

buildingstructuraldesign

intelligence”OR“generative”OR

“optimiz*”OR“explorat*”)ANDTS=

(“machinelearning”OR“deeplearning”

OR“neuralnetwork”OR“generativeadversarialnetwork”OR“variational

74

AI-basedbuildingstructuraldesign

autoencoder”OR“transformer”OR“diffusionmodel”))

Thesumofthesetwoformulas

188

methodsanddatasetconstructionusedinexistingresearch.

3.1.Datafeaturerepresentation

Table4-1

and

Table4-2

summarizethetensor-basedandgraph-baseddatarepresentationmethodsusedinrelatedresearch,and

Fig.5

illustratesthetypicalrepresentationmethods.Tensor-baseddatarep-resentationismorecommonlyusedincurrentresearch,whereasgraph-basedrepresentationsarelesscommon.Tensorrepresentationismoreintuitiveandstraightforward,andthusitismorewidelyusedincurrentgenerativeAI-basedintelligentdesignmethods.Althoughgraphsarebettersuitedforexpressingtopologicalfeatures,theyrequirecomplexandsophisticateddesignstorepresentpatternandsizefeatures.

(1)Tensor-basedrepresentation

(2)Graph-basedrepresentation

3.2.Datasetconstruction

Afterdeterminingthedatafeaturerepresentationmethod,thecollecteddatawereprocessedtoconstructthedataset.

Table5

providesdetailedinformationonthedatasetsdescribedinthecurrentliterature,includingthedesignobject,datacontent,dataquantity,andwhethertheyareopensource.Currently,theconstructionofdatasetsinbuildingstructuralintelligentdesignisrelativelyscarce,andtheintroductionofdatasetsisnotsufficientlycomprehensive.Onesignificantreasonisthelimitedavailabilityofpubliclyaccessiblefielddata.

Dataaugmentationmethodsarewidelyusedtoaddresstheissueoflimiteddata.(1)Imagedataaugmentationmethods,suchasflipping,symmetry,andtranslationofimages,caneffectivelyincreasethedatauptofourtimestheoriginaldatawithoutchangingtheattributesofthestructuraldesign[

13

,

104

].Augmentingdatathroughoverallimagesegmentationcanincreasethedatabyhundredsoftimesbypartitioningthecompleteddesignimageintoseverallocaldesignimageswithoutchangingthebuildingandstructuralsize

[71

,76

].(2)Tensordataaugmentationcangeneratehundredsorthousandsofdatabyparametricdesignforstructuraltopologyandcomponentsize[

30

,46

,72

].(3)Graphdataaugmentationcanalsobeperformedbasedonthespatialrelativecoordinatepositionineachnodeofthegraph,wheredatacanbeaugmented>3000timesbyoveralltranslation,flipping,androtatingcoordinates[

70

,74

].

3.3.Summaryofdatafeaturerepresentation

Currently,thereexistsignificantvariationsintheextractionandrepresentationofdatafeaturesfordifferentbuildingstructureforms.Whetheritinvolvespixel

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论