Page 111 - 2024年第55卷第8期
P. 111

interlayerlocationsandthicknesses.Theresultsshowthattheparticlesizehasasignificanteffectonthemigration -
                  depositionprocessoffineparticlesinsandcolumnswithinterlayers.Whentheparticlesizeratioislargerthan1 ,
                  thepeakvalueofthesuspendedparticlespenetrationcurveishigher,andtheamountoffineparticlesdepositedin
                  theinterlayerincreasesabruptly;theclosertheinterlayerpositionistothewaterinjectionend,thesmallerthe
                  thicknessis ,andthelargerthepeakvalueandthesuddenincreaseindepositionare.Whentheparticlesizeratio
                  islessthan1,thepeakvalueofthesuspendedparticlespenetrationcurveislower,andthereisnoparticleenrich
                  mentattheinterlayer;theclosertheinterlayeristothewateroutletend,thegreaterthethickness,andthehigher
                  thepeakvalueofthepenetrationcurve.Duringthetest,themigrationanddepositionofsuspendedparticleswill
                  causechangesinthepermeabilitycoefficient.Theprocessofpermeabilitycoefficientchangecorrespondstothe
                  changeofpenetrationcurve ,andtheprocessisdividedintothreekinds:deposition,strippingmigrationanddy
                  namicequilibrium.Fineparticlesaredepositedinsandysoilbyadsorptionandaccumulateattheadsorption.
                  Keywords:interlayer;particlesizeratio;migration;deposition;permeabilitycoefficient


                                                                                    (责任编辑:李 娜)


              (上接第 976页)
                       Researchonmulti - outputseismictime - historylong - term sequencesprediction
                                    modelforfreefieldofdam basedonvirtualsensors
                              1
                                                                                                 1
                                            2
                                                                                1
                                                                        1
                                                          3
                        SUZhe,LIUZongxian,YUHongling,TONGDawei,YUJia,WANGXiaoling
                 (1.StateKeyLaboratoryofHydraulicEngineeringIntelligentConstructionandOperation,TianjinUniversity,Tianjin 300072,China;
                                 2.YalongRiverHydropowerDevelopmentCompany,Ltd,Chengdu 610051,China;
                         3.CollegeofWaterResourcesandCivilEngineering,ChinaAgriculturalUniversity,Beijing 100083,China)
                  Abstract:Themultidimensionallong - termpredictionofseismictime - historyindam areasholdssignificantim
                  portanceforrapiddamageanalysis.Virtualsensors ,ascomplementarysensingmechanismstoseismicphysical
                  sensors ,facilitateseismictime - historypredictions.However,existingvirtualsensorsfacechallengesineffec
                  tivelypredictinglong - term sequencesformultiplesignals ,leadingtodelaysinanalyzingdam seismicdamage.
                  Addressingtheaforementionedissue ,amulti - outputseismictime - historylong - termsequencespredictionmodel
                  basedonTFA - Seq2Seqvirtualsensorsisproposed.ThismodelenhancestheSeq2Seqvirtualsensorsusingmulti -
                  tasklearning ,restructuringthemintoan“Encoder - 3Decoder”architecture.Thisstructureestablishesthemap
                  pingrelationshipbetweenmultipledamphysicalsensorsignalsandlong - termseismictime - historyinthreefree -
                  fielddirections.Additionally ,anattentionmechanismisintegratedtocapturetemporaldependenciesamongmul
                  tipleinputsignals ,resolvingsynchronousmulti - outputpredictionissuesandenhancingpredictionaccuracy.Fur
                  thermore ,Time - Frequencytransform(TF)layersandtheirinversetransformationlayersareintroducedtoimprove
                  theEncoderandDecoder ,shorteningthetemporallengthofseismicsignalsandextractingfrequencydomainfea
                  tures.Correspondingly ,amodeltrainingstrategyinvolvingstochasticforcedlearningisproposedtoovercomethe
                  limitationsofexistingvirtualsensorsineffectivelypredictinglongsequences.Casestudiesdemonstratethatthe
                  proposedmethodachievesavirtualsenseof10secondsaheadforseismicsignalsinthreedirectionswithindam
                  freefield.Comparedtomodelswithoutattentionmechanismsandsingleoutputs ,theproposedmethodexhibitsan
                  enhancedpredictionaccuracyof6.88% and3.32%,respectively.Thisresearchpresentsnovelinsightsandap
                  proachesforadvancingtheanticipatorysenseofseismicinformationduringseismicevents.
                  Keywords:freefieldseismic;virtualsensors;multi - outputlong - term sequencesprediction;TFA - Seq2Seq;
                  multi - tasklearning


                                                                                    (责任编辑:韩 昆)








                                                                                                —  9 8 9 —
   106   107   108   109   110   111   112   113   114   115   116