Efficient Protocol Design in Infrastructural RFID Systems

MISSING IMAGE

Material Information

Title:
Efficient Protocol Design in Infrastructural RFID Systems
Physical Description:
1 online resource (1 p.)
Language:
english
Creator:
Luo, Wen
Publisher:
University of Florida
Place of Publication:
Gainesville, Fla.
Publication Date:

Thesis/Dissertation Information

Degree:
Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Computer Engineering, Computer and Information Science and Engineering
Committee Chair:
CHEN,SHIGANG
Committee Co-Chair:
SAHNI,SARTAJ KUMAR
Committee Members:
PEIR,JIHKWON
HELMY,AHMED ABDELGHAFFAR
FANG,YUGUANG

Subjects

Subjects / Keywords:
energy-efficient -- rfid -- time-efficient -- trade-off
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre:
Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract:
RFID (radio-frequency identification) tags are becoming ubiquitously available in warehouse management, object tracking and inventory control. Researchers have been actively studying RFID systems as an emerging pervasive computing platform, which helps create a multi-billion dollar market. Small RFID tags, each with a unique ID, are attached to objects, allowing an RFID reader to quickly access the properties of each individual object or collect statistical information about a large group of objects. Much of the existing work on RFID systems is to design tag identification protocols that read the IDs from tags. Other work designs efficient protocols to estimate the number of tags in a large RFID system, detects missing tags, or collects useful information. The recent research trend is to improve time efficiency when designing an RFID solution because it is likely to be performed frequently. This is fine for passive tags - the kind used by Wal-Mart, which rely on radio waves emitted from an RFID reader to power their circuit and transmit information back through backscattering. Passive tags have short operational ranges, typically a few meters indoor, which seriously limits their applicability. For an application that covers a large area and involves a large number of tags, battery-powered active tags may be preferred for longer transmission ranges and richer on-tag hardware for more sophisticated functions. When battery is used, energy efficiency becomes a primary concern because it determines the lifetime of these tags. As an example, it may be ok to replace active tags once a couple years, but it will become a serious burden if they have to be replaced weekly, due to protocol designs that do not take energy efficiency into consideration. Hence, it is also necessary to study the methods of designing energy-efficient RFID solutions and controlling the performance tradeoff between an application's execution time and its energy cost. In this dissertation, we present two important and interesting problems, RFID missing tag detection and multigroup threshold-based classification. The former is used to detect the missing tag event within a pre-configured confidence interval, with respect to energy efficiency and time efficiency. The later focuses on using the minimum time to classify all above-threshold groups with a prescribed accuracy requirement.
General Note:
In the series University of Florida Digital Collections.
General Note:
Includes vita.
Bibliography:
Includes bibliographical references.
Source of Description:
Description based on online resource; title from PDF title page.
Source of Description:
This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility:
by Wen Luo.
Thesis:
Thesis (Ph.D.)--University of Florida, 2014.
Local:
Adviser: CHEN,SHIGANG.
Local:
Co-adviser: SAHNI,SARTAJ KUMAR.

Record Information

Source Institution:
UFRGP
Rights Management:
Applicable rights reserved.
Classification:
lcc - LD1780 2014
System ID:
UFE0046702:00001


This item is only available as the following downloads:


Full Text

PAGE 1

EFFICIENTPROTOCOLDESIGNININFRASTRUCTURALRFIDSYSTEMSByWENLUOADISSERTATIONPRESENTEDTOTHEGRADUATESCHOOLOFTHEUNIVERSITYOFFLORIDAINPARTIALFULFILLMENTOFTHEREQUIREMENTSFORTHEDEGREEOFDOCTOROFPHILOSOPHYUNIVERSITYOFFLORIDA2014

PAGE 2

c2014WenLuo 2

PAGE 3

Tomyparents,mywifeandmyson 3

PAGE 4

ACKNOWLEDGMENTS Firstofall,Iwouldliketothankmyadvisor,Prof.ShigangChen,forhisgreatsupport,guidanceandunderstandingthroughoutmygraduatestudy.Heisanincredibleadviser,apassionatescientist,andaterricperson.Withouthisconsistentsupportandencouragementthathelpedmetoovercomemanycrisissituations,theworkpresentedinthisdissertationwouldneverhaveexisted.Foreverythingyouhavedoneforme,Prof.Chen,Iwanttosaythankyoufromthebottomofmyheart.MyspecialthanksgotoProf.SartajSahni,Prof.YuguangFang,Prof.AhmedHelmy,Prof.Jih-KwonPeirandProf.YeXiafortheiradviceandsupportduringmystudyatUniversityofFlorida.IwouldalsoliketothankallthemembersinProf.Chen'sgroupfortheirhelp.TheyareMingZhang,YanQiao,ZhenMo,YianZhou,MinChen,andespeciallyforTaoLi,whooffersmealotofsuggestionsandencouragements.Finally,andmostimportantly,Iamthankfultoeachmemberofmyfamily:myparents,myparents-in-law,mybrother,andespeciallymywifeBobbiQiuliLaiandmysonPeterHaoyuLuo.Thankyouallforyoursupport,understanding,encouragement,andloveforsomanyyears. 4

PAGE 5

TABLEOFCONTENTS page ACKNOWLEDGMENTS .................................. 4 LISTOFTABLES ...................................... 8 LISTOFFIGURES ..................................... 9 ABSTRACT ......................................... 11 CHAPTER 1INTRODUCTION ................................... 13 1.1RFIDTechnologies ............................... 13 1.2ReviewofExistingResearchinRFIDSystems ............... 15 1.3ScopeofResearch ............................... 20 1.4Missing-TagDetectionandEnergy-TimeTradeoffinLarge-ScaleRFIDSystemsWithUnreliableChannels ...................... 21 1.5AnEfcientProtocolforRFIDMultigroupThreshold-BasedClassication 24 1.6OutlineofTheDissertation .......................... 25 2SYSTEMMODELSANDPERFORMANCEMETRICS .............. 27 2.1SystemModels ................................. 27 2.2PerformanceMetricsinRFIDSystems .................... 29 2.2.1PerformanceMetric1:ProtocolExecutionTime ........... 29 2.2.2PerformanceMetric2:EnergyCost .................. 30 2.2.3PerformanceMetric3:StorageCapacity ............... 30 2.2.4PerformanceMetric4:ComputationComplexity ........... 31 3RELATEDWORK .................................. 32 3.1RelatedWork1:RFIDAnti-CollisionProtocols ............... 32 3.2RelatedWork2:TrustedReaderProtocol(TRP) .............. 33 3.3RelatedWork3:CardinalityEstimationProtocolsandGroupTesting ... 34 4MISSING-TAGDETECTIONANDENERGY-TIMETRADEOFFINLARGE-SCALERFIDSYSTEMSWITHUNRELIABLECHANNELS ................ 37 4.1Preliminaries .................................. 38 4.1.1SystemModel .............................. 38 4.1.2Missing-TagDetectionProblem .................... 39 4.1.3PerformanceMetrics .......................... 40 4.1.4ClockSynchronization ......................... 41 4.1.5PriorWork ................................ 42 4.2AnIntermediateProtocol ........................... 43 4.2.1ProtocolDescription .......................... 43 5

PAGE 6

4.2.2Limitations ................................ 44 4.3EfcientMissing-TagDetectionProtocol(EMD) ............... 45 4.3.1ProtocolDesign ............................. 45 4.3.2ProbabilityofDetectingaMissing-TagEvent ............. 46 4.3.3Energy-TimeTradeoffCurve ...................... 49 4.3.4MinimumEnergyCost ......................... 51 4.3.5MinimumExecutionTime ....................... 52 4.3.6Energy-TimeTradeoff,TRP,andOfineComputation ........ 53 4.3.7FurtherEnergyReduction ....................... 54 4.4Multi-SeedMissing-TagDetectionProtocol(MSMD) ............ 55 4.4.1Motivation ................................ 55 4.4.2HashFunction ............................. 57 4.4.3BasicIdea ................................ 58 4.4.4Segmentation .............................. 59 4.4.5ProtocolOverview ........................... 59 4.4.6PhaseOne:TagAssignment ..................... 60 4.4.6.1Determiningsampledtags ................. 61 4.4.6.2Assigningsampledtagstosub-frames ........... 61 4.4.6.3Determiningseed-selectionsegments ........... 61 4.4.7PhaseTwo:Missing-TagDetection .................. 62 4.5Energy-TimeTradeoffinProtocolConguration ............... 63 4.5.1ExecutionTimeandEnergyCost ................... 64 4.5.2DetectionProbability .......................... 65 4.5.3Energy-TimeTradeoffCurve ...................... 67 4.5.4MinimumEnergyCost ......................... 68 4.5.5MinimumExecutionTime ....................... 69 4.5.6OfineComputation .......................... 70 4.5.7ConstrainedLeast-Time(orLeast-Energy)Problem ......... 70 4.5.8Impactofk ............................... 71 4.6MSMDoverUnreliableChannels ....................... 72 4.6.1MSMDunderRandomErrorModel .................. 74 4.6.2MSMDunderBurstErrorModel .................... 74 4.7NumericalResults ............................... 77 4.7.1Energy-TimeTradeoff ......................... 79 4.7.2PerformanceComparison ....................... 81 4.8Summary .................................... 81 5ANEFFICIENTPROTOCOLFORRFIDMULTIGROUPTHRESHOLD-BASEDCLASSIFICATION .................................. 84 5.1ProblemDenitionandSystemModel .................... 84 5.1.1SystemModel .............................. 84 5.1.2MultigroupThreshold-BasedClassicationProblem ......... 85 5.2Preliminary ................................... 86 5.2.1PriorWork ................................ 86 6

PAGE 7

5.2.2Motivation ................................ 89 5.3AnEfcientThreshold-BasedClassicationProtocol ............ 91 5.3.1DynamicSlotSharing ......................... 91 5.3.2Overview ................................ 92 5.3.3FramePhase .............................. 93 5.3.4ReportPhase .............................. 94 5.3.5Parameter-PrecomputingPhase ................... 96 5.4NumericalResults ............................... 99 5.4.1Setting .................................. 99 5.4.2ExecutionTimeRequiredwithRespectto1,1andl=h ...... 102 5.4.3FPRandFNRwithRespectto1,1andl=h ............ 103 5.4.4ExecutionTimeRequiredwithRespecttoAbove-ThresholdGroups 106 5.4.5ExecutionTimeRequiredwithRespecttoTheNumberofGroups 107 5.5Summary .................................... 108 6IMPLEMENTATION ................................. 109 6.1SystemSetup .................................. 109 6.2ASimpleMissing-tagDetectionPorotocolImplementation ......... 110 6.3ASimpleMulti-groupClassicationPorotocolImplementation ....... 112 6.4What'stobeexpected? ............................ 112 7CONCLUSIONANDPROPOSALFORFUTUREWORK ............. 114 7.1Conclusions ................................... 114 7.2FutureWork ................................... 115 REFERENCES ....................................... 117 BIOGRAPHICALSKETCH ................................ 122 7

PAGE 8

LISTOFTABLES Table page 4-1Notations ....................................... 40 4-2Theimpactofcorructedslots ............................ 73 4-3RelativeenergycostandexecutiontimeofMSMD(k=7)underpoptandpt,when=95%andn=50,000 ........................... 81 5-1Notations ....................................... 87 5-2Estimationtimecomparisonwhen1=99.9%and1=0.1% .......... 104 5-3Estimationtimecomparisonwhen1=99%and1=1% ............ 104 5-4Estimationtimecomparisonwhen1=95%and1=5% ............ 104 5-5Estimationtimecomparisonwhen1=90%and1=10% ........... 104 5-6FalsenegativeRatioandFalsePositiveRatiowhen1=99%and1=1% .. 105 5-7FalsenegativeRatioandFalsePositiveRatiowhen1=95%and1=5% .. 105 8

PAGE 9

LISTOFFIGURES Figure page 4-1DetectionprobabilityPemd(p,f)withrespecttotheframesizefwhenn=50,000,m=100,andp=5%. ........................... 49 4-2Framesizef(p)withrespecttosamplingprobabilitypwhenn=50,000,m=75,and=95%. ................................ 51 4-3Executiontimef(p)tswithrespecttoenergycostnp. ........... 51 4-4Energy-timetradeoffcurveintherangep2[popt,pt]. ............... 51 4-5Thevalueofptwithrespectton. .......................... 53 4-6Thevalueofpoptwithrespectton. ......................... 53 4-7InPhasetwo,thereaderbroadcaststheseed-selectionsegments,V1throughVf=l,oneatatime.EachsegmentViisimmediatelyfollowedbyasub-frameFioflslots,duringwhichthetagstransmit. .................... 59 4-8Arrowsrepresentthemappingfromtagstoslotsbasedonhashfunctions.Amongthem,thickarrowsrepresenttheassignmentoftagstoslots.Inthisexample,k=2. .................................... 60 4-9DetectionprobabilityPmsmd(p,f)withrespecttotheframesizefwhenn=50,000,m=100,k=3,andp=5%. ....................... 67 4-10Energy-timetradeoffcurve,i.e.,framesizef(p)withrespecttosamplingprobabilityp,whenn=50,000,m=75,k=3,and=95%. ................ 68 4-11Energy-timetradeoffcurveintherangep2[popt,pt],whichcorrespondstothecurvesegmenttotheleftofthedashedlineinFig. 4-10 ........... 69 4-12Thevalueofptwithrespectton. .......................... 69 4-13Thevalueofpoptwithrespectton. ......................... 69 4-14Energy-timetradeoffcurvesofEMDandMSMDunderdifferentkvalues,whenn=50,000,m=100,andp=5%. ......................... 72 4-15Protocolexecutiontimewithrespecttosamplingprobability,when=95%,m=50,andn=50,000. .............................. 80 4-16Zoom-inviewofenergy-timetradeoffinFigure 4-15 inthesamplingprobabilityrangeof[0,0.2]. ................................... 80 4-17Energycostandexecutiontimecomparisonwhenm=50and=99.9%. .. 82 4-18SameasthecaptionofFig. 4-17 exceptfor=99%. .............. 82 9

PAGE 10

4-19SameasthecaptionofFig. 4-17 exceptfor=90%. .............. 82 4-20Energycostandexecutiontimecomparisonwhenm=25and=99%. .... 82 4-21SameasthecaptionofFig. 4-20 exceptform=100. ............... 82 5-1TheestimationtimewithrespecttothegroupsizeforUPE,EZB,EnhancedFNEB,EMLEAandART,when01=99%and01=1%. ............. 87 5-2ExecutiontimewithrespecttothenumberofgroupsSwhicharesupposedtobereportedwhen1=99%,1=1%andl=0.7h.Thetotalnumberoftagsn1isxedtobe500,000ateachpoint. ...................... 103 5-3ExecutiontimewithrespecttothenumberofgroupsSwhicharesupposedtobereportedwhen1=99%,1=1%andl=0.7h.Thetotalnumberoftagsn1forallthegroupsincreasesalongwithS. .................... 103 5-4Executiontimewithrespecttothenumberofgroupswhen1=99%,1=1%,h=250,l=0.5h,andn1increaseswiththenumberofgroups. ....... 105 5-5Executiontimewithrespecttothenumberofgroupswhen1=99%,1=1%,h=250,l=0.5h,andn1isxedat500000. ................. 105 5-6Zoom-inofFigure 5-5 forTBCandGT. ....................... 106 6-1RFIDSystem ..................................... 110 6-2Timeframelength=100slots. ........................... 110 6-3Timeframelength=30slots. ............................ 111 6-4BeforevetagsareMssing. ............................. 111 6-5AftervetagsareMissing. ............................. 111 6-6Thepopulationestimationoffourgroups. ..................... 112 10

PAGE 11

AbstractofDissertationPresentedtotheGraduateSchooloftheUniversityofFloridainPartialFulllmentoftheRequirementsfortheDegreeofDoctorofPhilosophyEFFICIENTPROTOCOLDESIGNININFRASTRUCTURALRFIDSYSTEMSByWenLuoMay2014Chair:ShigangChenMajor:ComputerEngineeringRFID(radio-frequencyidentication)tagsarebecomingubiquitouslyavailableinwarehousemanagement,objecttrackingandinventorycontrol.ResearchershavebeenactivelystudyingRFIDsystemsasanemergingpervasivecomputingplatform,whichhelpscreateamulti-billiondollarmarket.SmallRFIDtags,eachwithauniqueID,areattachedtoobjects,allowinganRFIDreadertoquicklyaccessthepropertiesofeachindividualobjectorcollectstatisticalinformationaboutalargegroupofobjects.MuchoftheexistingworkonRFIDsystemsistodesigntagidenticationprotocolsthatreadtheIDsfromtags.OtherworkdesignsefcientprotocolstoestimatethenumberoftagsinalargeRFIDsystem,detectsmissingtags,orcollectsusefulinformation.TherecentresearchtrendistoimprovetimeefciencywhendesigninganRFIDsolutionbecauseitislikelytobeperformedfrequently.Thisisneforpassivetags-thekindusedbyWal-Mart,whichrelyonradiowavesemittedfromanRFIDreadertopowertheircircuitandtransmitinformationbackthroughbackscattering.Passivetagshaveshortoperationalranges,typicallyafewmetersindoor,whichseriouslylimitstheirapplicability.Foranapplicationthatcoversalargeareaandinvolvesalargenumberoftags,battery-poweredactivetagsmaybepreferredforlongertransmissionrangesandricheron-taghardwareformoresophisticatedfunctions.Whenbatteryisused,energyefciencybecomesaprimaryconcernbecauseitdeterminesthelifetimeofthesetags.Asanexample,itmaybeoktoreplaceactivetagsonceacoupleyears, 11

PAGE 12

butitwillbecomeaseriousburdeniftheyhavetobereplacedweekly,duetoprotocoldesignsthatdonottakeenergyefciencyintoconsideration.Hence,itisalsonecessarytostudythemethodsofdesigningenergy-efcientRFIDsolutionsandcontrollingtheperformancetradeoffbetweenanapplication'sexecutiontimeanditsenergycost.Inthisdissertation,wepresenttwoimportantandinterestingproblems,RFIDmissingtagdetectionandmultigroupthreshold-basedclassication.Theformerisusedtodetectthemissingtageventwithinapreconguredcondenceinterval,withrespecttoenergyefciencyandtimeefciency.Thelaterfocusesonusingtheminimumtimetoclassifyallabove-thresholdgroupswithaprescribedaccuracyrequirement. 12

PAGE 13

CHAPTER1INTRODUCTION 1.1RFIDTechnologiesBarcode[ 3 ]broughtarevolutionintheretailindustrybyspeedingupthecheckoutprocesswithalaserscannerthatreadstheproductIDfrombarcodeprintedoneachitemandretrievesitspriceautomaticallyfromadatabase.However,thereisaseriouslimitingfactor:barcodecanonlybereadinaverycloserangewithdirectsight,whichmakesitimpossibletobatch-accessobjectsthatarepiledonstoreracksorinshoppingcarts.RFID(radio-frequencyidentication)technologiesremovethislimitationbyintegratingsimplecommunication/storage/computationcapacitiesinattachabletags,whoseIDscanbereadwirelesslyoveradistance,evenwhenobstaclesexistbetweentagsandtheRFIDreader.TheRFIDtechnologiespromisetorevolutionizefutureinventorymanagement[ 35 45 52 ].Itisusedtotag,identifyandtrackindividualitems,casesandpalletsastheymovefromthemanufacturingoorthroughthesupplychainandintothehandsofthebuyerorconsumer.Astheobjectsmovethroughthesupplychain,wirelessRFIDreaderscancommunicatewithanRFIDtagontheobject,collectinformationabouttheobject(suchasauniquenumber)andmatchthatnumberinadatabasetoaccessacompleterecordabouttheobject.Thisreal-timetechnologyprovidesunprecedentedspeedandaccuracyinthesupplychain.Forexample,startingfromAugust1,2010,Wal-MarthasbeguntoembedRFIDtagsinclothing[ 50 ].Ifsuccessful,thesetagswillberolledoutontootherproductlinesatWal-Mar'smorethan3,750U.S.stores[ 5 ].Thatisonesteptowardscashier-lesscheckout,whereacustomerpusheshershoppingcarttopassanRFIDreaderatthecheckout,whereinformationintheembeddedtagsisautomaticallyreadandareceiptisprintedout.Today,practicalRFIDsystemsexistforautomatictollpayment,accesscontroltoparkinggarages,objecttracking,theftprevention,etc. 13

PAGE 14

RFIDtagsareattractivebecauseoftheirsimplicity,whichplacesthetechnicalchallengeinRFIDresearch.WecannotndanyspecicationonhowsimplefutureRFIDtagsshouldbe,butitisalwayspreferredthatasolutioncanachievecomparableefciencywithlesshardwarerequirement-thesimplerandcheaper,thebetter.Simplicityplacesconstraintonthesolutionspace,oftenmakinganotherwiseeasyproblemdifculttosolve.Forexample,informationcollectionisnotdifcultinaclassicalwirelessnetwork,whereeachclientimplementsrouting/scheduling/MACprotocols[ 10 11 ].IftheMACprotocolisCSMA/CA,theclientswillbeabletosensethechannelandtransmittheirinformationwhenitisidle.Inaddition,theyareabletodetectcollisionanduserandombackofftoresolveit.Now,supposewewanttouseRFIDtags-whicharedeployedonproductsinalargewarehouse,equipmentinahospital,orcarplatesinacity-forinformationcollection.WhatifthehardwareoftagsdoesnotsupportsuchanMACprotocol,letalonerouting/schedulingprotocols?Whatiftheirsimpleantennacannotsenseweaksignalfrompeersforcollisionavoidance,letaloneperformingrandombackoff?Yet,howdowecollectinformationfromalltagswiththebesttimeandenergyefciency?Thatbecomesachallenge.Thesametokengoesforvariousotherproblemsthatthisdissertationwillinvestigate.ThebasictechnologiesforRFIDhavebeenaroundforalongtime.Itsrootcanbetracedbacktoanespionagedevicedesignedin1945byLeonThereminoftheSovietUnion,whichretransmittedincidentradiowavesmodulatedwithaudioinformation.Inthepast,muchresearchconcentratedontwofronts:rst,physical-layertechnologiesfortransmittingIDsfromtagstoanRFIDreadermorereliably,overalongerdistance,andusinglessenergy;second,MAC-layertechnologiesforimprovingtherateatwhichareadercancollectIDsfromtags.Arguably,thesetechnologieswouldbesufcientiftheonlyrequirementfromapplicationswastoreadtagIDsasefcientlyaspossible.However,thefutureofRFIDwillgofarbeyondthis. 14

PAGE 15

Inrecentyears,arelativelysmallnumberofresearchgroups,havebeeninvestigatingnovelwaysinwhichfutureRFIDsystemscanbeusedtosolvepracticalproblems[ 20 22 25 27 28 31 37 40 48 51 53 55 58 ].Ofcourse,RFIDtagsmaybeembeddedinlibrarybooks,passports,driverlicenses,carplates,medicalproducts,etc.Inthecurrentapplicationmodel,tagsaretreatedasIDcarriersandtheyaredealtwithindividuallyforthepurposeofidentifyingtheobjectthateachtagisattachedto.Now,ifwemakeaparadigmshiftfromthisindividualviewtoacollectiveview,anarrayofnewapplicationsandinterestingresearchproblemswillemerge.Consideramajordistributioncenterofalargeretailer,assumingitappliesRFIDtagstoallitsmerchandise.Thesetags,whicharepervasivelydeployedinthecenter,shouldnotbetreatedjustasIDcarriersforindividualobjects.Collectively,theyconstituteanewinfrastructure,whichcanbeexploitedforcenter-wideapplications.Ifwetakeonestepfurther,wecanmakethisinfrastructuremorevaluablebyaugmentingtagswithminiaturizedsensors[ 19 ],suchthattheyreportnotonlystaticIDinformationbutalsodynamicreal-timeinformationabouttheirenvironmentorconditionsofthetagsthemselves.Ifwetakeanothersteptoconsidersecurityortagmobility,moreapplicationsandresearchproblemsopenup.Whilethesearebeyondtoday'sRFIDsystems,researchshouldtakethepioneerroletoexplorethepossibilitiesandestablishthefundamentalmethodsforwhatarefeasible. 1.2ReviewofExistingResearchinRFIDSystemsBelowwepresentarichsetofinterestingresearchestobeexploredintheeldofRFIDtechnologies.Weclassifytheseresearchesintodifferentcategories.TherstcategoryisSizeMeasurement.Thebasicproblemistoestimatethesizeofasystem,i.e.,thenumberoftags(orthenumberoftaggedobjects)inthesystem.Thesolutiontothisproblemcanbeusedtomeasuretheinventorylevelinawarehouseifgoodsaretagged,orestimatethenumberofpassengersinasubwaytrainifthesubwayticketsareembeddedwithRFIDtags.Acloselyrelatedproblemistoprecisely 15

PAGE 16

determine,insteadofestimating,thenumberoftags.Itisamoredifcultoperationandwilltakemoretimetocomplete.Next,supposeobjectsinalargeRFIDsystemareorganizedintogroups.Forinstance,inashoestore,eachgroupmaycorrespondtoaspecickindofshoesfromaspecicvendorwithaspecicproperty,e.g.,certainsize,type,orprice.Thegroupmeasurementproblemistoestimatethenumberoftags(i.e.,objects)ineachgroup.Consideranarbitrarygroup.Letkbetheactualnumberoftagsinthegroupand^kbetheestimatednumber.Wehavethefollowingaccuracyrequirement:Theprobabilityforktobeintherange[(1)]TJ /F4 11.955 Tf 12.65 0 Td[()^k,(1+)^k]isatleast,whereandaretwosystemparameters.Asimplerproblemistoidentifythegroupswhosesizesarebeyondathreshold.Wemayextendthisproblemwithmultiplethresholds,whichessentiallyspecifyasequenceofconsecutiveranges.Theproblembecomesclassifyingthegroupsintodifferentrangesbasedontheirsizes.Inaplacewhereobjectsmaybemovedinorout,sizemeasurementshouldbeperformedperiodically.Betweentwoconsecutivemeasurements,wemaywanttoknowhowmanynewobjectsaremovedin,howmanyexistingobjectsaremovedout,andhowmanyobjectsstayinthewarehousefortheperiodbetweenthemeasurements.Forexample,supposetwoconsecutivemeasurementresultsare10,000and20,000tags,respectively.Canweconcludethat10,000newobjectsaremovedin?No.Infact,thenumberofnewobjectsmaybeanynumberxbetween10,000and20,000,andthenumberofmoved-outobjectsisx)]TJ /F6 11.955 Tf 12.24 0 Td[(10,000.Itisadifcultproblemtondoutxwithoutkeepingtrackofindividualtags.ThesecondoneisAnomalyDetection.AninterestingapplicationofRFIDtagsistodetectmissingitemsinalargestorage.Consideramajorwarehousethatkeepsthousandsofapparel,shoes,pallets,cases,appliances,electronics,etc.Howtondoutifanythingismissing?Wemayhavesomeonewalkthroughthewarehouseandcountitems.Thisisnotonlylaboriousbutalsoerror-prone,consideringthatclothes 16

PAGE 17

maybestackedtogether,goodsonracksmayneedaladdertoaccess,andtheymaybeblockedbehindcolumns.IfweattachanRFIDtagtoeachitem,thewholedetectionprocesscanbeautomatedwithoneormultipleRFIDreaderscommunicatingwithtagstondoutwhetheranytags(andtheirassociatedobjects)areabsent.Theproblemofmissing-tagdetectionhastwoversions:exactdetectionandprobabilisticdetection.Theformeristoidentifyexactlywhichtagsaremissing.Thelatteristodetectamissing-tageventwithacertainpredenedprobability.Exactdetectiongivesmuchstrongerresults,butitsoverheadwillbegreaterthanprobabilisticdetection.Theoppositeofmissing-tagdetectionisunknown-tagdetection,wherewewanttoknowwhetherunexpectedtags(theirassociatedobjects)arepresentinthesystem-forinstance,aluggageinanairportisaccidentallymisplacedinawrongluggagegroup.Similarly,ithastwoversions:exactdetection,whichndsouttheexactIDsoftheunknowntags,andprobabilisticdetection,whichdetectsthepresenceofunknowntagswithacertainpre-denedprobability.Whenbothmissingtagsandunknowntagsoccur,wehavethemixedversionsoftheseproblems.Inaddition,thereisarelatedproblemcalledmisplacementdetection,whichdetectsobjectsmisplacedfromtheiroriginalplacementtoadifferentlocation.Thatmayoccurinaretailstorewhereacustomerpicksupanitem,decideslatertonotbuyit,anddropsitoffatadifferentplace.Next,InformationCollection.TheutilityofRFIDsystemswillbetremendouslyexpandedifminiaturizedsensorsareincorporatedintotags'circuit,enablingthemtocollectusefulinformationinrealtime.Asensormaybedesignedtomonitorthestateofthetagitself,forinstance,theresidualenergyofthebattery.Inanotherexample,consideralargechilledfoodstoragefacility,whereeachfooditemisattachedwithanRFIDtagthatcarriesathermalsensor.AnRFIDreadermayperiodicallycollecttemperaturereadingsfromtagstocheckwhetheranyareaistoohot(ortoocold),whichmaycausefoodspoil(orenergywaste).Theproblemishowtoefcientlycollect 17

PAGE 18

sensor-produceddatafromalargenumberoftagstoanRFIDreader,undertheconstraintofsimplecommunicationmodeladoptedbyRFIDsystems;seetheSystemModelinthissectionfordetails.Amoregeneralproblemistocollectinformationfromasubsetofalltags.Considerthepreviousexampleofachilledfoodstorage.Becauseeachareainthefacilitymaybepackedwithmanyfooditems,thetemperaturereadingsfromtheseclose-bytagsarehighlyredundant.Hence,itisnotnecessaryforthereadertocollectinformationfromalltagsinthesystem.Thereadermayselectasubsetoftagseachtimetocollecttemperaturedata.Itisaharderproblemtocollectinformationfromasubsetoftagsthanfromalltagsbecausethereaderhastomakesurethattagsthatarenotunderquerydonottransmit-theirtransmissionswillinterferewiththetransmissionsmadebytagsofinterest,particularlywhentheformeroutnumbersthelatterbyfar.Mobilitymayhelpinformationcollection.ConsiderthescenariowherecarplatesareintegratedwithsimpleRFIDtagsthatcanrecordalocationvalueandatimevaluetransmittedfromanRFIDreaderandthentransmitthembacktoanotherreaderatalatertime.SupposeRFIDreadersaredeployedatchosenlocationsinacityandtheirclocksarewellsynchronized.Inordertoprovidereal-timemeasurementonthelatencyfortravelingfromonelocationxtootherlocations,thereaderatlocationxmaybroadcastitslocationandthecurrenttime,whichwillberecordedbythepassingcars.Whenthesecarspassotherlocations,thereaderstherewillrequestforthatinformationandlearnhowlongittakesacartotravelfromlocationxtotheirplaces.Ifeachtagcanonlycarryoneoralimitednumberoflocation-timevalues,theproblemishowtocoordinateamongstthereaderssothattheyknowwhichpassingcarstheyshouldwritetheirvaluesandwhichcarstheyshouldonlyreadvaluesfrom,inordertoachievepre-speciedcity-wideaccuracyrequirementsonlatencymeasurement.ThefourthoneisPrivacyPreservation.Privacy-preservingauthenticationisaveryimportantproblem.InsecureRFIDsystems,areaderwillacceptatag'sinformation 18

PAGE 19

onlyafteritauthenticatesthetag.Atagmaybeattachedtoapharmaceuticalproductthatcarriespatientinformation,apassportthatcarriesaperson'sidentication,oracommercialproductthatcarriesinformationaboutmanufactureddate,expiringdate,ingredients,etc.Someapplicationsrequireprivacy-preservingauthentication,inwhichatagshouldnotgiveoutanyidentifyinginformationduringauthenticationprocess.SupposeapolicetriestouseamobileRFIDreadertoauthenticateadriver'slicenseembeddedwithanRFIDtag,andthereaderhasaccesstoadatabaseofallsecretkeysthatarepre-installedindriverlicenses.Thereaderhastoknowwhichkeyitshouldusetoperformauthentication.Inatypicalauthenticationprotocol,thelicensetransmitsanidentifyingnumbertothereader,whichwillusethatnumbertosearchthedatabasefortherightkey.However,thisleadstoasecurityloophole.Theidentifyingnumber,transmittedwirelessly,revealsthepresenceofthecarrier.Fakereadersmayinitiatetheauthenticationprocessatchosenlocationsandchosentimes.Theywillterminatetheprocessaftertheidentifyingnumberisreceived.Itallowsthemtomonitorthewhereaboutofthelicense'scarrier.Hence,aresearchproblemistodesignanauthenticationprotocolthatallowsanRFIDreadertoauthenticateatagwithoutrequiringthetagtotransmitanyidentifyingdata.Ideally,theinformationtransmittedfromataglookstotallydifferentandrandomeachtimethetagisauthenticated.Next,wemovetoprivacy-preservinginformationcollection.WehavediscussedthatfuturecarplatesmaybeintegratedwithRFIDtagstoenableinformationcollectionfrommovingvehicles.IfRFIDreadersaredeployedatintersections,theycancollectthetagIDsofpassingvehicles.FromthesecollectedIDs,wewillknownotonlypointtrafcvolumesatintersectionsbutalsopoint-to-pointvolumesbetweenanytwolocations-theremustavehicletravelingbetweenthemifbothlocationsrecordacommontagID.However,thisleadstoseriousprivacybreaching:Foreachregisteredvehicle,asitsIDisrecordedateachintersectionthatitpasses,itsentiremovinghistoryisrevealed,whichisagainsttheanonymitybydesignprincipleforprivacyprotectionrequired 19

PAGE 20

byIntelliDrive,formerlyknownasVehicleInfrastructureIntegrationorVII,aninitiativefromUSDepartmentofTransportation(USDOT).Hence,thechallengeistoallowthecollectionofstatisticalpoint-to-pointinformation,yetprotectinformationabouteachindividualvehicle. 1.3ScopeofResearchInprevioussection,wehaveshownthattherearenumerousproblemsandsolutionsinRFIDsystems,outofwhichtwoproblemsareveryinteresting,relativelynewandunderexploration,andhencehaveagreatpotentialforimprovement.Thesetwoproblemsaredescribedasfollows:First,considerawarehousethatstoresalargenumberofcommercialproductsoramilitarybasethatstockpilesalargequantityofgunsandammunitionpackages.Supposeeachobject(e.g.,amicrowaveoven,arie,orabulletmagazine)isattachedwithaRFIDtag,whichisabletocommunicatewithRFIDreaders.Asobjectsmoveinandoutofthewarehouse,asmall-rangereaderatadoorwillautomaticallycollecttheIDsofthenewtagsthatbecomepresentandremovetheIDsofthetagsthataremovedout[ 44 ].Now,ifsomeobjectsarestolenfromthewarehouse,howtotimelydetectsuchevents?Withoutanyautomatictools,wehavetoresorttomanualinventorywalk-through,whichislaborious,expensive,andslow.Suchoperationscannotbeperformedfrequently,andhencewillnothelpusdetectthetheftintimeinordertocatchthethieves.However,ifRFIDtagsareinstalled,themissing-tagdetectioncanbeautomatedandperformedfrequentlysothatanymajortheftisswiftlyreported.Therearetwodifferentmissing-tagdetectionproblems:exactdetectionandprobabilisticdetection.Theobjectiveofexactdetectionistoidentifyexactlywhichtagsaremissing.Theobjectiveofprobabilisticdetectionistodetectamissing-tageventwithacertainpredenedprobability.Anexactdetectionprotocol[ 24 42 44 56 ]givesmuchstrongerresults,butitsoverheadisfargreaterthanaprobabilisticdetectionprotocol[ 15 29 36 46 ].Hence,theybothhavetheirvalues.Infact,theyarecomplementary 20

PAGE 21

toeachother,andshouldbeusedtogether.Forexample,aprobabilisticdetectionprotocolmaybescheduledtoexecutefrequently,e.g.,onceeveryminute,inordertotimelycatchanylosseventsuchastheft.Onceitdetectssometagsaremissing,itmayinvokeanexactdetectionprotocoltopinpointwhichtagsaremissing.Ifoneexecutionofaprobabilisticdetectionprotocoldetectsamissing-tageventwith99%probability,veindependentexecutionswilldetecttheeventwith99.99999999%probability.Ifthatisnotenough,wemayscheduleanexactdetectionprotocol[ 24 ]everyvetimestheprobabilisticdetectionprotocolisexecuted.1.Inthisdissertation,weonlyconsiderprobabilisticdetection.Next,Weinvestigateadifferentproblem.Inpractice,tagsareoftenattachedtoobjectsbelongingtodifferentgroups,forinstance,differentbrandsofshoesinalargeshoestore,differenttitlesofbooksinabookstore,andgoodsfromdifferentcountriesormanufacturersinaport.Onechallengeistodeterminewhetherthenumberoftagsineachgroupisaboveorbelowaprescribedthresholdvalue.Thethresholdmaybesethightoidentifythepopulousgroups,itmaybesettoalevelthattriggerscertainactionssuchasreplenishingthestocks,orevenmultiplethresholdscanbeusedtoclassifygroupsbasedontherangeoftheirpopulationsizes.Solvingthismultigroupthreshold-basedclassicationproblemgivesusabasictooltoaccessalargepopulationofnumerousgroups. 1.4Missing-TagDetectionandEnergy-TimeTradeoffinLarge-ScaleRFIDSystemsWithUnreliableChannelsThisworkfocusesonaparticularlychallengingproblem,thedetectionofmissingtageventinlarge-scaleRFIDsystem.Thebasicdetectionmethodisintroducedinthepioneerwork[ 46 ]:ARFIDreadermonitorsatimeframeoffslots.Throughahash 1Anotherexactdetectionprotocol[ 44 ]reportswhichtagsaremissing,butdoesnotguaranteetoreportallmissingones.Theprotocolin[ 24 ]guarantees100%reporting. 21

PAGE 22

function,eachtagpseudo-randomlyselectsaslotinthetimeframetotransmit.Thereadercanpredictinwhichsloteachknowntagwilltransmit.Itdetectsamissing-tageventifnotagtransmitsduringaslotwhenthereissupposedtobetag(s)transmitting.However,multipletagsmayselectthesameslottotransmit.Ifatagismissing,itsslotmaybekeptbusybytransmissionfromanothertag.Consequently,thereadercannotguaranteethedetectionofamissing-tagevent.Furthermore,weobservethattimeefciencyshouldnotbethesoleperformanceconsiderationwhendesigningamissing-tagdetectionprotocol.Theenergycostmaybeanevenmorecriticalconcernifactivetagsareused.Duetolimitedoperationaldistances,passivetagsaremostlyusedforsmall-rangeapplicationssuchasfastcheckout.Forfutureautomaticinventorymanagementfunctionsthatcoveraverylargearea,activetagsarelikelytobethechoice.Activetagsusetheirownpowertotransmit.Alongerrangecanbeachievedbytransmittingatahigherpower.Theyarealsoricherinresourcesforimplementingadvancedfunctions.Theirpricebecomeslessofaconcerniftheyareusedforexpensivemerchandizes(suchasrefrigerators)orreusedmanytimesasgoodsmovinginandoutofthewarehouse.Butactivetagsalsohaveaproblem.Theyarepoweredbybatteries.Rechargingbatteriesfortensofthousandsoftagsisalaboriousoperation,consideringthatthetaggedproductsmaystackup,makingtagsnoteasilyaccessible.Toprolongthetags'lifetimeandreducethefrequencyofbatteryrecharge,allfunctionsthatinvolvelarge-scaletransmissionbymanytagsshouldbemadeenergy-efcient.Theprotocolin[ 46 ]onlyconsiderstimeefciency,butnotenergyefciency.Afollow-upwork[ 29 ]furtherimprovesthetimeefciency.Firneret.al.[ 15 ]designasimplecommunicationprotocol,Uni-HB,todetectmissingitemsforfail-safepresenceassurancesystemsanddemonstrateitcanleadtolongersystemlifetimeandhighercommunicationreliabilitythanseveralpopularprotocols.Theprotocolhoweverdoesnotconsidertimeefciencyandrequiresalltagstoparticipateandtransmit,whichwillbelessefciencythanasampling-basedprotocol 22

PAGE 23

designthatrequiresonlyasmallfractionofthetagstoparticipate.Similarly,themethodin[ 36 ]alsorequiresalltagstoparticipate.Thepriorworkhasstudiedenergy-efcientprotocolsforestimatingthenumberoftagsinaRFIDsystem[ 26 ],orenergy-efcientanti-collisionprotocolsthatminimizetheenergyconsumptionofamobilereaderwhenthereaderisusedtocollecttheIDsofthetags[ 20 34 ].Webelievethisdissertationisthersttostudyenergy-efcientsolutionsforthemissing-tagdetectionproblem.Inaddition,muchoftheexistingliteratureassumesthatthecommunicationchannelbetweenaRFIDreaderandtagsisreliable,whichmeansthatinformationtransmittedisnevercorrupted.However,inreality,errorsmayoccurduetolowsignalstrength,noise,orinterferenceintheenvironment.Theoccurrenceoferrorsusuallyfollowsacertaindistribution,whichischaracterizedbyanerrormodel,describingthestatisticalpropertiesofunderlyingerrorsequences.Tosolvethemissingtagdetectionproblem,rst,weproposetwonew,moresophisticatedprotocoldesignsformissing-tagdetection.Theytakebothenergyefciencyandtimeefciencyintoconsideration.TherstoneiscalledEMD,whichworksinasimilarwaytoTheprotocolin[ 46 ],exceptthattagsaresampledwithasmallprobability.Inthesecondprotocol,byintroducingmultiplehashseeds,ournewdesignprovidesmultipledegreesoffreedomfortagstochoosewhichslotstheywilltransmitin.Thisdesigndrasticallyreducesthechanceofcollision,andconsequentlyachievesmultiple-foldreductioninbothenergycostandexecutiontime.Insomecases,thereductionismorethananorderofmagnitude.Second,withthenewdesigns,werevealafundamentalenergy-timetradeoffinmissing-tagdetection.Ouranalysisshowsthatbetterenergyefciencycanbeachievedattheexpenseoflongerexecutiontime,andviceversa.Theperformancetradeoffcanbeeasilycontrolledbyacoupleofsystemparameters.Throughouranalyticalframeworkforenergy-timetradeoff,weareabletocomputetheoptimalparametersettingsthatachievethesmallestprotocol 23

PAGE 24

executiontimeorthesmallestenergycost.Theframeworkalsoenablesustosolvetheenergy-constrainedleast-timeproblemandthetime-constrainedleast-energyprobleminmissing-tagdetection.Third,weextendourprotocoldesigntoconsiderchannelerrorundertwodifferenterrormodels.Ourprotocolscanbeconguredtoworkundertheseerrorconditions. 1.5AnEfcientProtocolforRFIDMultigroupThreshold-BasedClassicationMostexistingworkadoptsaatRFIDsystemmodelandperformsfunctionsofcollectingtagIDs,estimatingthenumberoftags,ordetectingthemissingtags.However,inpractice,tagsareoftenattachedtoobjectsofdifferentgroups,whichmayrepresentadifferentproducttypeinawarehouse,adifferentbookcategoryinalibrary,etc.Aninterestingproblem,calledmultigroupthreshold-basedclassication,istodeterminewhetherthenumberofobjectsineachgroupisaboveorbelowaprescribedthresholdvalue.Solvingthisproblemisimportantforinventorytrackingapplications.Preciseclassicationrequiresustoknowtheprecisenumberoftagsineachgroup.Tagidenticationprotocols[ 8 12 17 20 32 34 49 54 ]candothat,butittakesthemsignicanttimetocompleteifthenumberoftagsisverylarge.Onewaytoimproveefciencyisrelaxingtheproblemfromaccurateclassicationtoapproximateclassication[ 45 ],wheretheclassicationaccuracycanbetunedtomeetapre-denedrequirement.Wemayusecardinalityestimationprotocols[ 21 22 37 49 54 ]toestimatethenumberoftagsineachgroup,andclassifythegroupbasedontheestimation.However,thoseprotocolsareefcientwhenestimatingasmallnumberoflargegroups,buttheyarenotefcientwhenestimatingalargenumberofsmallgroups,becausetheirexecutiontimeforeachgroupislargelyindifferentingroupsize,aswewilldemonstrateshortly.In[ 45 ],Shengetal.applygrouptestingtoapproximatelydetectpopulargroups.Whenthenumberofgroupsabovethethresholdissmall,theirperformanceisgood.However,theperformanceofthegroup-testing-basedsolution 24

PAGE 25

degradesquickly(intermsoftheexecutiontime)whenthenumberofgroupsabovethresholdbecomeslarge.Inthisdissertation,weproposeanewclassicationprotocolthatisscalabletoalargenumberofgroups.Itsdesignisdrasticallydifferentfromtraditionalapproachesthatmeasurethesizeofonegroupatatime.Itmeasuresthesizesofallgroupstogetheratonceinamixedfashion.Yet,thenewprotocolisabletoperformthreshold-basedclassicationwithanaccuracythatcanbepre-settoanydesirablelevel,allowingtradeoffbetweentimeefciencyandaccuracy.Ourmaincontributionsaresummarizedasfollows:First,wedesignanewprotocolforthreshold-basedclassicationinamulti-groupRFIDsystembasedontagsamplingandlogicalbitmaps,whichsharetimeslotsuniformlyatrandomamongallgroupsduringtheprocessofmeasuringtheirpopulations.Weusethemaximumlikelihoodestimationmethodtoextractper-groupinformationfromthesharedslots.Suchslotsharinggreatlyreducestheamountoftimeittakestocompleteclassication.Samplingfurtherimprovetheperformanceoftheprotocolsignicantly.Futhermore,givenanaccuracyrequirement,weshowanalyticallyhowtocomputeoptimalsystemparametersthatminimizetheprotocolexecutiontimeundertheconstraintoftherequirement.Ourestimationmethodbasedonsamplingandlogicalbitmapsensuresthatfalsepositive/falsenegativeratiosarebounded,wherefalsepositiveoccurswhenabelow-thresholdgroupisreportedasabove-thresholdandfalsenegativeoccurswhenanabove-thresholdgroupisnotreported.Finally,wecomprehensivelyevaluatetheproposedsolutionandcompareitwithexistingprotocols.Oursimulationresultsmatchwellwiththeanalyticalresults,whichdemonstratethatthenewprotocolperformsfarbetterintermsofexecutiontimethanthebestexistingwork. 25

PAGE 26

1.6OutlineofTheDissertationTherestofthedissertationisorganizedasfollows:Chapter2presentsthesystemmodelsandvariousperformancemetricsconsideredinRFIDsystem.Chapter3istherelatedworkofMissingtagdetectionandRFIDmultigroupthreshold-basedclassication.Chapter4presentsanRFIDmissing-tagdetectionprotocolinlarge-scaleRFIDSystemswithunreliablechannels,andrevealafundamentalenergy-timetradeoff.Inthissection,weproposeanovelprotocoldesignthatconsidersbothenergyefciencyandtimeefciency.Itachievesmulti-foldreductioninbothenergycostandexecutiontimewhencomparingwiththebestexistingwork.Chapter5proposesaniterativeprotocolforthreshold-basedclassicationinamulti-groupRFIDsystembasedonlogicalbitmapsthatsharetimeslotsuniformlyatrandomamongallgroupsduringtheprocessofmeasuringtheirpopulations.Chapter6presentstheimplementationofthesimpleversionofourprotocols.Chapter7drawstheconclusionsandbrieydiscussesthepotentialfuturework. 26

PAGE 27

CHAPTER2SYSTEMMODELSANDPERFORMANCEMETRICS 2.1SystemModelsTherearethreetypesofRFIDtags.Passivetagsaremostwidelydeployedtoday.Theyarecheap,butdonothaveinternalpowersources.PassivetagsrelyonradiowavesemittedfromanRFIDreadertopowertheircircuitandtransmitinformationbacktothereaderthroughbackscattering.Theyhaveshortoperationalranges,typicallyafewmetersinanindoorenvironment.Tocoveralargearea,arraysofRFIDreaderantennasmustbeinstalled.Semi-passivetagscarrybatteriestopowertheircircuit,butstillrelyonbackscatteringtotransmitinformation.Activetagsusetheirownbatterypowertotransmit,andconsequentlydonotneedanyenergysupplyfromthereader.Activetagsoperateatamuchlongerdistance,makingthemparticularlysuitableforapplicationsthatcoveralargearea,whereoneorafewRFIDreadersareinstalledtoaccessalltaggedobjectsandperformmanagementfunctionsautomatically.Withricheronboardresources,activetagsarelikelytogainmorepopularityinthefuture,particularlywhentheirpricesdropovertimeasmanufacturaltechnologiesareimprovedandmarketsareexpanded.Communicationsbetweenthereaderandthetagsaretime-slotted.Thereader'ssignalwillsynchronizetheclocksofthetags.Insomeprotocols,thecommunicationisdrivenbythereaderinarequest-and-responsepattern.Thereaderissuesarequest,whichisfollowedbyatimeframeconsistingoffslots,duringwhichthetagsmayrespondbytransmittingsomeinformation.Aslotissaidtobeemptyifnotagresponds(transmits)intheslot.Itiscalledasingletonslotifexactlyonetagresponds.Itisacollisionslotifmorethanonetagresponds.Asingletonorcollisionslotisalsocalledabusyslot.TheEPCglobalGen-2standard[ 14 ]requiresaslotlengthof10bitsinordertodistinguishsingletonslotsfromcollisionslots.Wedenethelengthoflong-responseslotasTlongsothathtereader 27

PAGE 28

cantellwhetherthereiscollisioninaslot.Onthecontrary,onebitisenoughifweonlyneedtodistinguishemptyslotsfrombusyslots`0'meansemptyand`1'meansbusy.Hence,tagresponseswillbemuchshorter(orconsumemuchlessenergy)ifaprotocolonlyneedstoknowempty/busyslots,liketheoneinthisdissertationdoes.AframetakesfTshorttime,whereTshortisthetimeofaslot.Whenfislarge,thetimeittakestheRFIDreadertotransmititsrequest,whichisasmallconstant,canbeignored.Fortimeefciency,weshouldminimizetheframesizef,subjecttothedetectionrequirementinSection 4.1.2 .Forenergyefciency,weshouldminimizethetotalnumberofresponsesfromalltags.Becausethereareemptyslotsandcollisionslots,thenumberofresponsesisnotthesameasthenumberoftimeslotsintheframe.Aswewillseelater,reducingthenumberofresponsesmayrequireustoincreasethenumberoftimeslotsinordertomeetthedetectionrequirement.LetTtagbethetimeittakestotransmitatagID.Obviously,Ttag>TshortbecauseittakesalongertimetotransmitmultiplebitsinanIDthanone-bitinformationinatagresponse.BasedonthespecicationoftheEPCglobalGen-2standard[ 14 ],wedeterminethatTtag=2609.76sfora96-bittagIDandTshort=290.81s,aftertherequiredwaitingtimes(e.g.,gapofidletimebetweentransmissions)areincluded.Signicantasymmetryexistsbetweenreadersandtags:Unliketags,thecostforareaderisnotaseriousconcernbecauseitisnotneededinlargequantities;oftentimes,onlyoneisneeded.Therefore,unliketags,thereaderisnotlimitedinstoragespace,computationpower,orenergysupply.Ifnecessary,itcanbeconnectedtoapowerfulserverforresources.Withahigh-qualityantenna,areaderisabletoreceiveweaksignalsfromtags.Withlow-qualityantennas,althoughtagscanreceivestrongsignalsfromthereader,theycannotreceiveeachother'sweaksignals.Theymaynotevensensewhetherthechannelisbusyoridle,i.e.,whetheranothertagistransmitting.Norcantheysenseifcollisionhasoccurredwhentwotagstransmitsimultaneously.Butthereadercandetectwhetherthechannelisidleorwhethercollisionoccurs.Such 28

PAGE 29

asymmetrypointsoutadesignprinciplethatweshouldfollow:pushingthecomplexitytothereaderwhileleavingthetagssimple. 2.2PerformanceMetricsinRFIDSystemsThereexistfourimportantperformancemetricsintheprotocoldesignofaRFIDsystem. 2.2.1PerformanceMetric1:ProtocolExecutionTimeImaginethatalargeretailerhasawarehouseinitsdistributioncenterthatregularlystorestensofthousandsofelectronics,furniture,apparel,shoes,pallets,cases,etc.Amissing-tagdetectionprotocolisexpectedtobeexecutedfrequently(e.g.,onceevery15minutes)inordertotimelyraisealarmsuponunexpectedremovalofobjectsfromthewarehouse.However,falsedetectionofmissingtagsmayoccurifnormaloperationsremoveobjectsduringthetimewhentheprotocolisexecuting.Inabusywarehouse,asgoodsconstantlymoveinandout,falsealarmsmayhappeneveniftheprotocol'sexecutiontimeisanumberofseconds.Hence,itishighlydesirablethattheexecutiontimeiskeptassmallaspossibleinordertominimizethedisruptiontonormalinventoryoperations.(Ourworkisabletoreducetheexecutiontimeto1secondorlessundercertainparametersettings.Insuchasmallduration,itisextremelyunlikelythatsometagsareremovedbynormaloperationsthataremechanicalinnature.)Theexecutiontimeisaffectedbytwomajorfactors.Therstfactorishowstringentthesystemrequirementis.Aprotocolsuchas[ 46 ]willtakefarmoretimetodetectonemissingtagwith99.9%probabilitythantodetect100missingtagswith95%.Hence,inordertocontroltheprotocol'sexecutiontime,apracticalsystemmaybeconguredwithm=100and=95%.Itmeansthatasingleprotocolexecutionwilldetectthemissing-tageventwith95%ifm=100.Becausetheprotocolisexecutedperiodically,aftertheithexecution,thedetectionprobabilitybecomes1)]TJ /F6 11.955 Tf 12.25 0 Td[((1)]TJ /F6 11.955 Tf 12.25 0 Td[(95%)i,whichrapidlyapproachesto100%whenibecomeslarge.Evenifthenumberofmissingtagsissmallerthanmandthusthedetectionprobabilityofoneprotocolexecutionissmaller 29

PAGE 30

than,themissing-tageventwilleventuallybedetectedafterasufcientnumberofexecutions.Thesecondfactorthathasamajorimpactonexecutiontimeistheprotocoldesign.Inthisdissertation,weshowthatabetterprotocoldesigncanreducetheexecutiontimeconsiderablywithoutanysignicantincreaseinthecomplexityoftheonlineoperations. 2.2.2PerformanceMetric2:EnergyCostInordertosupportadvancedmanagementfunctionsthatcoveralargearea,battery-poweredactivetagsareabetterchoicebecausetheyhavemuchlongertransmissionranges.Ifpassivetagswereused,onewouldhavetotaketheRFIDreaderandmovearoundthewholearea,collectinginformationfromlocationtolocation,orelseadensereaderarrayhastobeinstalledtoextendthecoverage.Activetagsallowoneorafewreaderstocollectinformationfromalargearea.Whenactivetagsareused,wemustconservetheirbatterypowerinordertoprolongthetags'lifetimebeforetheyhavetoberecharged.Thetaggedgoods(suchasapparel)maystackinpiles,andtheremaybeobstacles,suchasrackslledwithmerchandize,betweenatagandthereader.Weexpecttheactivetagsaredesignedtotransmitwithsignicantpowerthatishighenoughtoensurereliableinformationdeliveryinsuchademandingenvironment.TheenergyconsumedbytheRFIDreaderisoflessconcernbecauseitsbatteriescanbeeasilyrechargedoritmayevenuseanexternalpowersource.Weassumethatthereadertransmitsatsufcientlyhighpower. 2.2.3PerformanceMetric3:StorageCapacityItisknownthataRFIDtagislimitedinmemory,from512bitsto128kilobytes[ 6 ],whichispreloadedwithatagIDandothernecessaryinformationtoidentifyitself,andhenceleaveslessmemoryforprotocolsinstalledinthetag.Tomakethematterevenmorechallenging,someprotocolsmayrequiretobeinitializedwithabunchofauthenticationkeysoragreatvolumeofmemorytostoretheintermediatedata.Inthatcase,thoseprotocolmayhavetobeinstalledinsomespecialtagswithlargememory 30

PAGE 31

thatismuchexpensive,whichseriouslylimitstheirapplicability.Hence,ourRFIDprotocoldesignshouldbeasmemory-efcientaspossible. 2.2.4PerformanceMetric4:ComputationComplexityThecomputationcomplexityisalsoaveryimportantfactorwithregardstocomputation-constraintRFIDtags.ThelimitationoflackingnecessarycomputationalresourcestosupportstrongcryptographicauthenticationschememakessecuringRFIDsystemaverychallengingtask.Therefore,whendesigningdifferentprotocols,weshouldtakegoodcareofthecomputationpartsuchthatthecomputationalpriceshouldbeloworthemajorcomputationworkliesonthereader. 31

PAGE 32

CHAPTER3RELATEDWORK 3.1RelatedWork1:RFIDAnti-CollisionProtocolsMostexistingworkonRFIDsystemsistodesignthetag-collectionprotocols,thatreadtheIDsfromthetags.Theymainlyfallintotwocategories.Oneistree-based[ 2 4 7 33 60 ]andtheotherisALOHA-based[ 9 23 41 47 49 57 ].Thetree-basedprotocolsorganizeallIDsinatreeofIDprexes[ 4 33 60 ].Eachin-treeprexhastwochildnodesthathaveoneadditionalbit,`0'or`1'.ThetagIDsareleavesofthetree.TheRFIDreaderwalksthroughthetree.Asitreachesanin-treenode,itqueriesfortagswiththeprexrepresentedbythenode.Whenmultipletagsmatchtheprex,theywillallrespondandcausecollision.Thenthereadermovestoachildnodebyextendingtheprexwithonemorebit.Ifzerooronetagresponds(intheone-tagcase,thereaderreceivesanID),itmovesupinthetreeandfollowsthenextbranch.Anothertypeoftree-basedprotocolstriestobalancethetreebylettingthetagsrandomlypickwhichbranchestheybelongto[ 2 7 33 ].Inthebestcase,whenthetreeiscompletelybalanced,thereadertakestwo(orthree,dependingontheactualdesign)queriesonaveragefordiscoveringeachID.However,itwillincurmuchmoreoverheadfortheaverageandworstcases.TheALOHA-basedprotocolsworkasfollows.Thereaderrstbroadcastsaqueryrequest.EachtagchoosesatimeslottotransmititsID.Ifatagselectsaslotthatnoneofothertagsselect,itcanbesuccessfullyidentiedandwillkeepsilentfortherestoftheprocess.Ifmultipletagstransmitsimultaneously,theresponsesaregarbledduetocollisionandretransmissionsarerequired.Theprocessterminateswhenallthetagsaresuccessfullyidentied.TheenhanceddynamicframedslottedALOHA(EDFSA)[ 23 ]increasestheidenticationprobabilitybyadjustingtheframesizeandrestrictingthenumberofrespondingtagsintheframe. 32

PAGE 33

3.2RelatedWork2:TrustedReaderProtocol(TRP)ThemostrelatedworkwendintheliteratureistheTrustedReaderProtocol(TRP)byTan,ShengandLi[ 46 ].Forsecurityreasons,theirsystemconsistsofaserverandaRFIDreader.TheformerstoresthetagIDsandperformsthecomputation,whilethelattercommunicateswiththetags.Weassumethereaderistrusted.Tosimplifytheprotocoldescription,welogicallycombinetheserverandthereaderintoasingleentity,stillcalledthereader.AdesigngoalofTRPistoreducethetimeofthedetectionprocess.Toinitiatetheexecutionoftheprotocol,aRFIDreaderbroadcastsadetectionreques,askingthetagstorespondinatimeframeoffslots.Thedetectionrequesthastwoparameters,theframesizefandarandomnumberr.EachtagmapsitselftoaslotintheframebyhashingitsIDandr.Itthentransmitsduringthatslot.Thereaderrecordswhichslotsarebusyandwhichareempty.Thisisbinaryinformationwhereeachslotcarrieseither`1'or`0'.Multiplereadersmaybeusedtoextendthecoverage.Inthiscase,allreaderswillsimultaneouslymonitorthetimeslotsintheframe.Aslotisconsideredtobebusyifanyreaderrecordsthattheslotisbusy.BecausethereaderknowstheIDsofalltags,itknowswhichtagsaremappedtowhichslots.Morespecically,itknowswhichslotsareexpectedtobesingletons,i.e.,oneandonlyonetagismappedtoeachofthem.Ifanexpectedsingletonslotturnsouttobeempty,thetagthatismappedtothisslotmustbemissing.Notalltagsaremappedtosingletonslots.Whentwoormoretagsaremappedtothesameslot(acollisionslot),ifonlyoneofthetagsismissing,theslotwillremainbusyandthusthemissing-tageventwillnotbedetected.Obviously,ifweincreasetheframesize,collisionwillbelesslikelyandtherewillbealargernumberofsingletonslots,whichmeanstheprobabilityforanytagtomaptoasingletonisgreater.Intheeventofmissingtags,theprobabilitythatanexpectedsingletonturnsouttobeemptyisalsogreater,andhencetheprobabilityofdetecting 33

PAGE 34

themissing-tageventisgreater.Therequirementisthatwemustdetectthemissing-tageventwithaprobabilityofatleastifmormoretagsaremissing.TRPisdesignedtominimizeitsexecutiontimebyusingthesmallestframesizethatensuresadetectionprobabilityof.AseriouslimitationofTRPisthatitonlyconsiderstimeefciency.Itisnotenergy-efcientbecausealltagsmustbeactiveandtransmitduringthetimeframe.B.Firneretal.[ 15 ]considerenergycost,buttheirprotocolrequiresalltagstoparticipateandtransmit,whichwillbelessefcientthanasampling-basedsolutionwhereonlyasmallfractionoftagsparticipate.Inthisdissertation,weshowTRParespecialcasesofamuchbroaderprotocoldesignspace.NotonlyarethereprotocolcongurationsthatperformmuchbetterthanTRPintermsofbothtimeandenergyefciencies,butwealsorevealafundamentalenergy-timetradeoffinthisdesignspace,whichallowsustoadaptprotocolperformancetosuitvariousneedsinpracticalsystems.Therecentworkin[ 24 ]alsorequiresalltagstotransmit,andittakesmuchlongertimethanTRP.However,itsolvesadifferentproblembyexactlyidentifyingwhichtagsaremissing.Hence,theprotocolin[ 24 ]complementsTRPandourworkinthisdissertation.ThecheaperprotocolofTRPorourprotocolcanbeexecutedfrequentlytoidentifythemissing-tagevent.Oncesuchaneventisdetected,theexpensiveprotocolin[ 24 ]canbeexecutedtondwhichtagsaremissing. 3.3RelatedWork3:CardinalityEstimationProtocolsandGroupTestingSheng,TanandListudiedthemultigroupthreshold-basedclassicationproblemin[ 45 ].Theybeginwithasimplethresholdcheckingscheme(TCS)toapproximatelyanswerswhetherthenumberoftagsexceedsathreshold.BasedonTCS,theyproposetwoprobabilisticprotocols.Therstoneisbasedongenericgrouptesting(GT),whichconsistsofmultiplerounds.Ineachround,thereadershufesallgroupsintodifferentcategories,eachofwhichmaycontaintagsfrommultiplegroups.TCSisthenapplied 34

PAGE 35

tocheckthenumberoftagsineachcategory.Thecategorieswithsufcienttagsarelabeledaspotentialpopulouscategories,whichmayincludeabove-thresholdgroups.Intheend,thetestinghistoryisusedtoclassifyallabove-thresholdgroups.Thesecondprotocolisacombinationofgrouptestinganddivide-and-conquer,whichignoresthecategoriesthatfailtopasstheTCStestsinthepreviousround,dividestheremainingcategoriesintomultiplesub-categories,andappliesTCStoeachsub-categoriesintheremainingrounds.Anotherpossiblesolutionforthemultigroupthreshold-basedclassicationproblemistouseareadertocollecttheactualtagIDsfromtags[ 8 12 17 20 32 34 49 54 ],whereeachIDcontainsbitsthatidentifythegroupofthetag.Applyingtotheprobleminthisdissertation,theseID-collectionprotocolsdonotworkwellforlarge-scaleRFIDsystemsduetotheirlongidenticationtime.ManymethodswereproposedtoestimatethewholepopulationofanRFIDsystem.Theyareessentiallysingle-groupestimators.Wecanusethemtorstestimateindividualgroupsizes(onegroupatatime)andthenusethesizesforclassicationpurpose.KodialamandNandagopalproposetherstsetofsingle-groupestimators,includingtheZeroEstimator(ZE),theCollisionEstimator(CE),andtheUniedProbabilisticEstimator(UPE),whichcollectinformationfromtagsinaseriesoftimeframesandestimatesthewholepopulationoftagsinthesystembasedonthenumberofemptyslotsand/orthenumberofcollisionslots[ 21 ].Afollow-upworkbythesameauthorsproposestheEnhancedZero-BasedEstimator(EZB)[ 22 ],whichisanasymptoticallyunbiasedestimatorandmakesestimationonlybasedonthenumberofemptyslots.Qianetal.provideareplicate-insensitiveestimationalgorithmcalledtheLottery-Framescheme(LoF)[ 37 ].TheEnhancedFirstNon-EmptyslotsBasedEstimator(EnhancedFNEB)[ 16 ]canbeusedtoestimatetagpopulationinbothstaticanddynamicenvironmentsbymeasuringthepositionoftherstnon-emptyslotineachframe.Lietal.[ 26 ]studytheestimationproblemforlarge-scaleRFID 35

PAGE 36

systemsfromtheenergyanglebasedonaEnhancedMaximumLikelihoodEstimationAlgorithm(EMLEA).Theydesignseveralenergy-efcientprobabilisticalgorithmsthatiterativelyreneacontrolparametertooptimizetheinformationcarriedinthetransmissionsfromtags,suchthatboththenumberandthesizeofthetransmissionsareminimized.TheAverageRunbasedTagestimation(ART)scheme[ 43 ]furtherreducestheexecutiontimeforpopulationestimation,basedontheaveragerunlengthofonesinthebitstringreceivedinthestandardizedframe-slottedAlohaprotocol.Finally,theZero-OneEstimator(ZOE)[ 59 ]providesfastandreliablecardinalitybytuningthesystemparametersandconvergingtotheoptimalsettingsthroughabisectionsearch. 36

PAGE 37

CHAPTER4MISSING-TAGDETECTIONANDENERGY-TIMETRADEOFFINLARGE-SCALERFIDSYSTEMSWITHUNRELIABLECHANNELSRFID(radiofrequencyidentication)technologiesarepoisedtorevolutionizeretail,warehouseandsupplychainmanagement.Oneoftheirinterestingapplicationsistoautomaticallydetectmissingtagsinalargestoragespace,whichmayhavetobeperformedfrequentlytocatchanymissingeventsuchastheftintime.BecauseRFIDsystemstypicallyworkunderlow-ratechannels,pastresearchhasfocusedonreducingexecutiontimeofadetectionprotocoltopreventexcessively-longprotocolexecutionfrominterferingnormalinventoryoperations.However,whenactivetagsareusedforalargespatialcoverage,energyefciencybecomescriticalinprolongingthelifetimeofthesebattery-poweredtags.Furthermore,muchofexistingliteratureassumesthatthechannelbetweenareaderandtagsisreliable,whichisnotalwaystrueinrealitybecauseofnoise/interferenceintheenvironment.Giventheseconcerns,thisdissertationmakesthreecontributions:First,weproposeanovelprotocoldesignthatconsidersbothenergyefciencyandtimeefciency.Itachievesmulti-foldreductioninbothenergycostandexecutiontimewhencomparingwiththebestexistingwork.Second,werevealafundamentalenergy-timetradeoffinmissing-tagdetection,whichcanbeexiblycontrolledthroughacoupleofsystemparametersinordertoachievedesirableperformance.Third,weextendourprotocoldesigntoconsiderchannelerrorundertwodifferentmodels.Wendthatenergy/timecostwillbehigherinunreliablechannelconditions,buttheenergy-timetradeoffrelationpersists.Therestofthischapterisorganizedasfollows:Section 4.1 givesthesystemmodelandproblemdenition,aswellasthepriorart.Section 4.2 presentsanintermediateprotocolthatcansolvethemissingtagdetectionproblem.Section 4.3 isamoreefcientprotocolcalledEMD.Section 4.4 proposesournalmissing-tagdetectionprotocol.Section 4.5 investigatesenergy-timetradeoffinprotocolconguration.Section 4.6 37

PAGE 38

extendstheprotocolundertwodifferenterrormodels.Section 4.7 evaluatestheprotocolthroughsimulations.Section 4.8 drawstheconclusion. 4.1Preliminaries 4.1.1SystemModelTherearethreetypesofRFIDtags.Passivetagsaremostwidelydeployedtoday.Theyarecheap,butdonothaveinternalpowersources.PassivetagsrelyonradiowavesemittedfromaRFIDreadertopowertheircircuitandtransmitinformationbacktothereaderthroughbackscattering.Theyhaveshortoperationalranges,typicallyafewmetersinanindoorenvironment,whichseriouslylimitstheirapplicability.Semi-passivetagscarrybatteriestopowertheircircuit,butstillrelyonbackscatteringtotransmitinformation.Activetagsusetheirownbatterypowertotransmit,andconsequentlydonotneedanyenergysupplyfromthereader.Activetagsoperateatamuchlongerdistance,makingthemparticularlysuitableforapplicationsthatcoveralargearea,whereoneorafewRFIDreadersareinstalledtoaccessalltaggedobjectsandperformmanagementfunctionsautomatically.Withricheron-boardresources,activetagsarelikelytogainmorepopularityinthefuture,whentheirpricesdropovertimeasmanufacturetechnologiesareimprovedandmarketsareexpanded.Theyareparticularlyattractiveforhigh-valuedobjectssuchasluxurybags,laptops,cellphones,TVs,etc.,orwhenthetagsarereusedoverandoveragain.Communicationbetweenareaderandtagsistime-slotted.Thereader'ssignalsynchronizestheclocksoftags.Therearedifferenttypesoftimeslots[ 24 ],amongwhichtwotypesareofinterestinthisdisseartation.Thersttypeiscalledatag-IDslot,whoselengthisdenotedasTtag,duringwhichareaderisabletobroadcastatagID.Thesecondtypeiscalledashort-responseslot,whoselengthisdenotedasTshort,duringwhichatagisabletotransmitone-bitinformationtothereader,forinstance,announcingitspresence. 38

PAGE 39

Inthisdissertation,weconsideractivetags.Besidecommunicatingwithareader,weassumethetagshavethefollowingcapability:performingahashfunction,carryingasmallinternalstoragetokeepafewparametersanda96-bitseed-selectionsegmentfromthereader,beingabletocheckthevaluesinthesegment,andhavingaclockthatenablesatagtotransmitataspecicslotofatimeframeorwaitupatapre-scheduledtime. 4.1.2Missing-TagDetectionProblemTheproblemistodesignanefcientprotocolforaRFIDreadertodetectwhethersometagsaremissing,subjecttoadetectionrequirement:Asingleexecutionoftheprotocolshoulddetectamissing-tageventwithprobabilityifmormoretagsaremissing,whereandmaretwosystemparameters.Forexample,considerabigshoestorethatcarriestensofthousandsofshoes,andwemaysettheparameterstobe=99%andm=10,sothatoneexecutionoftheprotocolwilldetectanyeventofmissing10ormoreshoeswith99%probability.Ifweperformindependentexecutionsoftheprotocolperiodically,thedetectionprobabilityofanymissingeventwillapproachto100%,nomatterwhatthevaluesofandmare.Furthermore,aswehaveexplainedintheintroduction,alow-overheadprobabilisticdetectionprotocolmaybeusedinconjunctionwithahigh-overheadexactdetectionprotocol(whichisscheduledmuchlessfrequently)tocatchanymiss.WeassumethattheRFIDreaderhasaccesstoadatabasethatstorestheIDsofalltags.Thisassumptionisnecessary[ 46 ].Withoutanypriorknowledgeofatag'sexistence,howcanweknowthatitismissing?TheassumptioncanbeeasilysatisedifthetagIDsarereadintoadatabasewhennewobjectsaremovedintothesystem,andtheyareremovedfromthedatabasewhentheobjectsaremovedoutthisiswhatatypicalinventorymanagementprocedurewilldo.Evenifsuchinformationislostduetoadatabasefailure,wecanrecovertheinformationbyexecutinganID-collectionprotocol[ 4 18 33 41 57 ]thatreadstheIDsfromthetags.Inthiscase,wewillnot 39

PAGE 40

Table4-1. Notations SymbolsDescriptions nnumberofRFIDtagsinthesystemmnumberofmissingtagsprobabilityofdetectingthemissing-tageventfnumberofslotsinatimeframerrandomnumberTtagtimefortransmittingatagIDTshorttimefortransmittingone-bitinformationknumberofhashseedsPmsmdprobabilityforMSMDtodetectamissing-tageventPemdprobabilityforEMDtodetectamissing-tageventpsamplingprobabilityinMSMDorEMDf(p)framesizethatminimizestheexecutiontimeunderagivensamplingprobabilitypfoptoptimalframesizethatachievesthesmallestexecutiontime,fopt=minff(p)gpptsamplingprobabilityunderwhichfoptisachieved,i.e.,fopt=f(pt)poptoptimalsamplingprobabilitythatminimizestheenergycost detectmissing-tageventsthathavealreadyhappened.However,oncewehavetheIDsoftheremainingtags,wecandetectthemissing-tageventsafterthispointoftime.Notations(mostofwhichareintroducedlater)aresummarizedinTable 4-1 forquickreference. 4.1.3PerformanceMetricsWeconsidertwoperformancemetrics,executiontimeoftheprotocolandenergycosttothetags.First,RFIDsystemsuselow-ratecommunicationchannels.Lowrates,coupledwithalargenumberoftags,oftentakeRFIDprotocolslongtimetonishtheiroperations.Hence,inordertoapplysuchprotocolsinabusywarehouseenvironment,itisdesirabletoadoptnoveldesignstoreduceexecutiontimeasmuchaspossible.Second,activetagscarrylimitedbatterypower.Replacingtagsisatedious,manualoperation.Onewayofsavingenergyistominimizethenumberoftagsthatareneededtoparticipateineachprotocolexecution.Whenatagparticipatesinaprotocolexecution,ithastopoweritscircuitduringtheexecution,receiverequestinformation 40

PAGE 41

fromthereader,andtransmitback.Whenatagdoesnotparticipate,itgoesintothesleepmodeandincursinsignicantenergyexpenditure.TheenergycosttotheRFIDreaderislessofaconcernbecausethereader'sbatterycanbeeasilyreplacedoritmaybepoweredbyanexternalsource. 4.1.4ClockSynchronizationForanymissing-tagdetectionprotocolthatisscheduledtoexecuteatxedtimeintervals,thereisaneedtosynchronizetheclocksofthetagssothattheycanwakeupattherightmoments.Toachieveverylowenergyconsumptionduringsleep,theseactiveRFIDtagsmayuselow-powerRCoscillatorsasclocksources,whichhoweverhaverelativelylargedrift.Thedriftwillbecomesignicantiftheclockisleftunsynchronizedforanextendedperiodoftime.Followingtheargumentintheintroduction,weexpectthemissing-tagdetectionprotocolstorunfrequently.Onesolutiontodealwiththedriftproblemistocalibratealltags'clocksatthebeginningofeachscheduledprotocolexecutioninordertokeepthemsynchronized.Thetagsthatarenotsupposedtoparticipateinaroundofexecutionwillgobacktosleepafterclocksynchronization.Itwilladdaxedamountofenergyexpendituretoallmissing-tagdetectionprotocolsthatrequirethereadertopullinformationfromtagsatxedtimeintervals.Becausesynchronizationoverheadiscommontoallsuchprotocols,wewillnotincludeitinperformancecomparison.Alsonotethatthisoverheadisrelativelysmallwhencomparingwiththeenergycostneededtopoweratagforreceiving,transmitting,andcomputingintheentiredurationofaprotocolexecution.Anothersolutionisonlylettingtheparticipatingtagswakeup,buttheyneedtowakeupalittleearlierthanscheduledtocompensatefortheclockdrift,suchthattheycanreceivetherequestsfromthereader.Thisapproachalsoincursadditionalenergyoverheadbecausetagshavetobepoweredalittlelongerforreceiving.Theenergyoverheadforclocksynchronizationdoesnotexistforapush-basedmissing-itemdetectionprotocolsuchasUni-HB[ 15 ],whereeverysensor(attachedto 41

PAGE 42

anitem)transmitsitsIDandasequencenumbertoabasestationineachepoch.Inordertolowerthecollisionprobability,Uni-HBspreadssensortransmissionsineachepoch,whichmeanstheprotocolexecutioncontinuesovereachepochsincetheIDsmaybesentbythesensorsatanytime.Intheintroduction,however,wehavearguedforshortprotocolexecutiontimetoavoidinterferencefrombusywarehouseoperationsthatmoveitemsinandout.Furthermore,Uni-HBrequiresalltagstoparticipatebysendingtheirIDstothebasestationineachepoch,whereaswepreferanapproachthatinvolvesonlyafractionoftagstosaveenergy.Forthesereasons,Uni-HBisnotsuitableformeetingtherequirementsinthisdissertation. 4.1.5PriorWorkWerstdescribetheTrustedReaderProtocol(TRP)byTan,ShengandLi[ 46 ].Givenatimeframeoffslots,theRFIDreadermapseachtagtoaslotintheframebyhashingitsIDandarandomnumberr.Afterthereadermapsalltagstotheslots,itclassiesslotsintothreecategories:Aslotissaidtobeemptyifnotagismappedtotheslot.Itiscalledasingletonslotifexactlyonetagismappedtotheslot.Itisacollisionslotifmorethanonetagismappedtotheslot.BecausethereaderknowstheIDsofalltags,itknowswhichtagsaremappedtowhichslots.Itknowsexactlywhichslotsareempty,whicharesingletons,andwhicharecollisionslots.Toinitiatetheexecutionoftheprotocol,aRFIDreaderbroadcastsadetectionrequest,askingthetagstorespondinatimeframeoffslots.Thedetectionrequesthastwoparameters,theframesizefandtherandomnumberr.Afterreceivingtherequest,eachtagmapsitselftoaslotintheframethroughthesamehashfunction.Itthentransmitsduringthatslot.Listeningtothechannel,thereaderrecordsthestateofeachslot,whichiseitherbusywhenoneormoretagstransmitoridlewhennotagtransmits.Thisisbinaryinformationwhereeachslotcarrieseither`1'or`0'.Whenatagtransmits,itdoesnothavetosendanyparticularinformation.Itonlyneedstomakethechannelbusy.When 42

PAGE 43

notagismissing,thereaderexpectsallsingletonandcollisionslotsarebusy.However,ifthereaderndsanexpectedbusyslottobeactuallyidle,itknowsthatthetag(s)thatismappedtothisslotmustbemissing.TRPisdesignedtominimizeexecutiontimebyusingthesmallestframesizethatensuresadetectionprobabilityifmormoretagsaremissing.Certainly,iffewertagsaremissing,thedetectionprobabilitywillbelower.Afollow-upwork[ 44 ]essentiallyexecutesTRPiterativelytoidentifywhichtagsaremissing.AseriouslimitationofTRPisthatitonlyconsiderstimeefciency.Itisnotenergy-efcientbecausealltagsmustbeactiveandtransmitduringthetimeframe.B.Firneretal.[ 15 ]considerenergycost,buttheirprotocolrequiresalltagstoparticipateandtransmit,whichwillbelessefcientthanasampling-basedsolutionwhereonlyasmallfractionoftagsparticipate.Inthisdissertation,weshowTRPisspecialcaseofamuchbroaderprotocoldesignspace.NotonlyarethereprotocolcongurationsthatperformmuchbetterthanTRPintermsofbothtimeandenergyefciencies,butwealsorevealafundamentalenergy-timetradeoffinthisdesignspace,whichallowsustoadaptprotocolperformancetosuitvariousneedsinpracticalsystems. 4.2AnIntermediateProtocolWepresentanenergy-efcientprotocolthatservesasanintermediatedesignsteptowardsournalsolutioninthenextsection. 4.2.1ProtocolDescriptionItiswellknownthatasmallgroupofpeoplemuchfewerthan365canhaveahighprobabilitytocontaintwowhocelebratetheirbirthdaysonthesameday.Thisiscalledthebirthdayparadox.Similarly,tworelativelysmallsubsetsofthetagsinalargeRFIDsystemcanalsohavealargeprobabilityofsharingacommontag.LetMbethesetofmmissingtags,andKbeasubsetoftagsthatthereaderrandomlyselectsfromtheinventorylistofntagscurrentlyinthesystem.Thereaderperformsasimple 43

PAGE 44

operationtoverifythepresenceofthesejKjtags,wherejKjisthenumberofthetagsinK.IttransmitstheIDsofthesetagsoneafteranother.AftertransmittinganID,thereaderwaitsforashortperiodandlistensforaresponse.WhenatagreceivesitsID,itwillacknowledgeitspresencebysendingaresponse.IfthereaderdoesnotreceiveanyresponsebackforanID,itreportsthemissingtagevent.TheideaisthatifjKjisreasonablylarge,KandMwillhaveagoodchancetoshareatleastonecommontag.Inotherwords,thereaderwillndthatthepresenceofatleastonetaginKcannotbepositivelyconrmed.Hence,themissing-tageventisdetected.Thisintermediateprotocolisenergy-efcientbecausethereareoveralljKjtagresponses,insteadofnresponsesrequiredbyTRP.NotethattheenergyconsumptionbythereaderfortransmittingtheIDsisasecondaryconcern. 4.2.2LimitationsWeobservethattheintermediateprotocolhasmuchroomforimprovement.First,itisnottime-efcient.AlthoughonlyasubsetoftagIDsisselected,ittakesaconsiderableamountoftimetoverifythepresenceofeachselectedtag.AccordingtotheparametersoftheEPCglobalGen-2standard[ 14 ],ittakesabout2609.76stotransmitatagIDof96bits,andittakes290.81sforatagtorespondwithonebitacknowledgmentafterreceivingitsID.Incomparison,ittakesjust290.81sforaRFIDreadertoidentifyeachemptyorbusyslotintheframeusedbyTRP.Second,althoughonlyasubsetoftagstransmitandeachofthemonlytransmitsonce,theyhavetolistentothechannelfortheirIDs,whichmeansthattheircircuitshavetobecontinuouslypoweredtoreceiveuptokIDs.Itistruethattransmittingislikelytoconsumemuchmorepowerthanreceivingifthesamenumberofbitsareinvolved.However,inourintermediateprotocol,eachselectedtagmakesjustoneshorttransmission,butithastoreceivealargenumberofbits,whichmakestheaggregatereceivingenergysignicant. 44

PAGE 45

Inournextprotocol,forbettertimeefciency,wemakesurethatnotagIDsaretransmitted.Forbetterenergyefciency,wemakesurethateachtagonlyneedstoreceiveatmostonepollingrequest. 4.3EfcientMissing-TagDetectionProtocol(EMD) 4.3.1ProtocolDesignWeproposeournalsolutionthataddressesthelimitationsoftheintermediateprotocolintheprevioussection.TheRFIDreaderinitiatestheprotocolexecutionbybroadcastingapollingrequest.Uponreceiptoftherequest,eachtagdecideswithasamplingprobabilitypwhethertoparticipateinthepolling.Ifitdecidestoparticipate,itwillrandomlyselectaslotinthesubsequentframetorespond.Ifitdecidesnottoparticipate,itwillsimplyenterthesleepmodeandwakeupatthenextscheduletimefortheprotocolexecution.Alldecisionsaremadepseudo-randomlyandpredictablebythereader.Werstexplainhowtoimplementsampling.Thepollingrequestconsistsofthreeparameters:aframesizef,arandomnumberr,andanintegerx=dpXe,whereXisalarge,pre-conguredconstant(e.g.,216).Afterreceivingtherequest,atagperformsahashH1(id,r),whereidisthetag'sIDandH1isahashfunctionwhoserangeis[0,X).IfH1(id,r)
PAGE 46

acertainnumberofbitscounter-clockwiseaftertherthbitinthering.Thetwohashfunctionsindependentlypicktheir0thbitsatrandompositionsinthering.ThenumberofbitsreturnedbyH1(id,r)islog2X.ThenumberofbitsreturnedbyH2(id,r)isdlog2fe,andthenalhashoutputisthereturnedvaluemodulof.TheRFIDreaderhastheIDsofalltags,fromwhichitcanderiveallinformationthatisneededtoknowwhichtagswillparticipateinthepollingand,ifso,atwhichslotstheywilltransmit.Hence,itknowsexactlywhichslotsintheframewillbeemptyandwhichareexpectedtobebusy.Attheendoftheframe,ifthereaderndsthataslotthatissupposedtobebusyturnsouttobeempty,itknowsthatthetag(s)thatisexpectedtotransmitintheslotmustbemissing.Inthiscase,thereaderreportsamissing-tagevent.Whenmultiplesynchronizedreadersareusedtoextendthecoverage,wetreataslotasbusyifanyreaderrecordsthattheslotisbusy.Clearly,theexpectednumberoftagsthatwilltransmittheirresponsesintheframeisnp.Inthefollowingsection,wewillproposeanothermechanismthatensuresonlythetagsthattransmitintheframehavetoreceivethepollingrequestfromthereader.Othertagsthatdonottransmitinthisroundofpollingwillnotevenhavetoreceivetherequest.Obviously,itrequiressomemodicationstotheprotocol.Wedeliberatelydelaythedescriptionofthismechanismbecauseitonlymakessenseafterwepresentthenecessaryanalyticalresult.Theonlineoperationoftheaboveprotocol,calledEMD(EfcientMissing-tagDetection),isverysimple,whichwebelieveisanadvantageforaRFIDsystem.Mostoftheprotocolcomplexitybelongstotheofineoperations,whichdeterminethevaluesofpandfandrevealafundamentalenergy-timetradeoffthathasnotbeeninvestigatedbefore. 4.3.2ProbabilityofDetectingaMissing-TagEventExcludingthemmissingtags,therearen)]TJ /F3 11.955 Tf 13.09 0 Td[(mtagsremaininginthesystem.Consideranarbitrarytag.Itdecideswhethertoparticipateinthepollingwithasampling 46

PAGE 47

probabilityp.Ifitdecidestoparticipate,itwillrandomlymapitselftoaslotintheframe.Toassistouranalysis,weseriallyapplythissampling-mappingoperationtothetagsoneafteranother.LetEibethenumberofemptyslotsremainingintheframeafteritagsareprocessed.Clearly,E0=fandE1=f)]TJ /F6 11.955 Tf 12.21 0 Td[(1.En)]TJ /F7 7.97 Tf 6.59 0 Td[(misthenumberofemptyslotsafteralln)]TJ /F3 11.955 Tf 12.47 0 Td[(mtagsareprocessed;itisthenumberofemptyslotsthattheRFIDreaderwillobserveattheendoftheframe.ConsiderEi,2in)]TJ /F3 11.955 Tf 12.26 0 Td[(m,asrandomvariables.Letebeaconstantintherangeof[0,f]andProbfEi=egbetheprobabilityforEi=e.WederivearecursiveformulatocomputethevalueofProbfEi=eg.TheeventEi=ehappensintwocases:i)whenEi)]TJ /F5 7.97 Tf 6.59 0 Td[(1=eandtheithtagisnotmappedtoanemptyslot,orii)whenEi)]TJ /F5 7.97 Tf 6.59 0 Td[(1=e+1andtheithtagismappedtoanemptyslot.Furthermore,whenEi)]TJ /F5 7.97 Tf 6.59 0 Td[(1=e,theconditionalprobabilityfortheithtagtonotmaptoanemptyslotis1)]TJ /F3 11.955 Tf 12.24 0 Td[(pe f.WhenEi)]TJ /F5 7.97 Tf 6.59 0 Td[(1=e+1,theconditionalprobabilityfortheithtagtomaptoanemptyslotispe+1 f.Hence,wehave ProbfEi=eg=ProbfEi)]TJ /F5 7.97 Tf 6.59 0 Td[(1=eg(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pe f)+ProbfEi)]TJ /F5 7.97 Tf 6.59 0 Td[(1=e+1gpe+1 fBasedontheaboverecursiveformula,weconstructadynamicprogrammingalgorithm(seeAlgorithm 1 )tocomputeProbfEn)]TJ /F7 7.97 Tf 6.59 0 Td[(m=eg,for0ef,whichisessentiallytheprobabilitydensityfunction(p.d.f.)ofEn)]TJ /F7 7.97 Tf 6.58 0 Td[(m.ThearrayP[e],0ef,isinitializedwiththevaluesofProbfE0=eg,whicharezerosfor0ef)]TJ /F6 11.955 Tf 12.71 0 Td[(1andonefore=f.EachiterationintheforloopoveriupdatesthearraywithvaluesofProbfEi=eg.Attheendoftheloop,thearrayP[e]storesProbfEn)]TJ /F7 7.97 Tf 6.58 0 Td[(m=eg.Next,weconsiderthenumberofmissingtagsthatwouldhaveparticipatedinthepollingiftheyhadnotbeenmissing.isarandomvariablethatfollowingaBinomialdistributionwithparametersmandpbecauseeachofthemmissingtagswouldhavea 47

PAGE 48

Algorithm1Computep.d.f.ofEn)]TJ /F7 7.97 Tf 6.58 0 Td[(m INPUT:n,m,p,f OUTPUT:ProbfEn)]TJ /F7 7.97 Tf 6.59 0 Td[(m=eg,for0ef P[e]=0for0ef)]TJ /F6 11.955 Tf 11.96 0 Td[(1,P[f]=1; P[f+1]=0;nnauxiliaryvariable fori=1ton)]TJ /F3 11.955 Tf 11.95 0 Td[(mdo fore=0tofdo P0[e]=P[e](1)]TJ /F3 11.955 Tf 11.96 0 Td[(pe f)+P[e+1]pe+1 f; endfor fore=0tofdo P[e]=P0[e]; endfor endfor returnP[e]for0ef; probabilityofptoparticipate.Hence,foranyj2[0,m],Probf=jg=mjpj(1)]TJ /F15 10.909 Tf 10.91 0 Td[(p)m)]TJ /F7 7.97 Tf 6.59 0 Td[(j. (4)BecausetheRFIDreaderhasalltheinformation,itknowsthesubsetoftagsinthesystemthatwilldecidetoparticipateinthepolling.Thissubsetmayincludeanumberjofmissingtagsthatwouldhaveparticipatediftheyhadnotbeenmissing,wherej2[0,m].Eachofthesejtagsisrandomlymappedtoaslot.Ifthisisanemptyslotobservedbythereaderattheendoftheframe,thereaderwilldetectthemissing-tageventbecauseitexpectsatagtotransmitatthisslot.Undertheconditionthatthereareeemptyslotsobservedbythereaderattheendoftheframe,theprobabilityforonemissingtagtomaptoanobservedemptyslotise f.Undertheconditionthattherearejmissingtagsthatwouldhaveparticipatedinthepolling,theprobabilityforanyofthemtomaptoanobservedemptyslotis1)]TJ /F6 11.955 Tf 12.19 0 Td[((1)]TJ /F7 7.97 Tf 13.4 4.71 Td[(e f)j.Whenthathappens,theRFIDreaderdetectsamissing-tagevent. 48

PAGE 49

Figure4-1. DetectionprobabilityPemd(p,f)withrespecttotheframesizefwhenn=50,000,m=100,andp=5%. Summarizingtheabovereasoning,wegivetheprobabilityforEMDtosuccessfullydetectamissing-tagevent,whichisafunctioninpandf,denotedasPemd(p,f).Pemd(p,f)=fXe=0ProbfEn)]TJ /F7 7.97 Tf 6.58 0 Td[(m=egmXj=0Probf=jg(1)]TJ /F17 10.909 Tf 10.91 0 Td[((1)]TJ /F15 10.909 Tf 12.15 7.38 Td[(e f)j)=fXe=0ProbfEn)]TJ /F7 7.97 Tf 6.58 0 Td[(m=egmXj=0mjpj(1)]TJ /F15 10.909 Tf 10.91 0 Td[(p)m)]TJ /F7 7.97 Tf 6.59 0 Td[(j(1)]TJ /F17 10.909 Tf 10.91 0 Td[((1)]TJ /F15 10.909 Tf 12.16 7.38 Td[(e f)j) (4)Therstsigmaintheformulasumsovertheconditionalprobabilitiesforthenumberofemptyslots.Thesecondsigmaintheformulasumsovertheconditionalprobabilitiesforthenumberofmissingtagsthatwouldhaveparticipatedinthepolling.ThevaluesofProbfEn)]TJ /F7 7.97 Tf 6.59 0 Td[(m=egarecomputedbyAlgorithm 1 4.3.3Energy-TimeTradeoffCurveAsmallervalueoffmeansashorterprotocolexecutiontime,whereasasmallervalueofpmeansasmallernumberoftransmittingtags,whichinturnmeansasmallerenergycost.Wecannotarbitrarilypickthevaluesoffandp.TheymustsatisfytherequirementPemd(p,f).Subjecttothisconstraint,weshowthatthevaluesoffandpcannotbeminimizedsimultaneously.Thechoiceoffandprepresentsanenergy-timetradeoff.Ifwexthevalueofp,Pemd(p,f)becomesafunctionoff.Itisanincreasingfunctionoffbecausealargerframeincreasestheprobabilityforamissingtagtomap 49

PAGE 50

toasingletonslot.Asthetagismissing,theexpectedsingletonbecomesanemptyslotobservedbythereader,resultinginmissing-tagdetection.ThesolidlineinFig. 4-9 showsanexampleofthecurvePemd(p,f)withrespecttofwhenn=50,000,m=100,andp=5%.BecausePemd(p,f)isanincreasingfunction,theminimumvalueoffthatsatisestherequirementPemd(p,f)canbefoundbysolvingtheequationPemd(p,f)=.Thesolutionisdenotedasf(seeFig. 4-9 foranillustration).Foreachdifferentsamplingprobabilityp,wecancomputethesmallestusableframesizeffromtheequationPemd(p,f)=.Hence,fcanbeconsideredasafunctionofp.f(p)=minffjPemd(p,f)^fUg, (4)whereUisanupperboundoftheframesize,asapracticalRFIDsystemmayconsideraframesizebeyondacertainupperboundtobeunacceptableduetoexcessivelylongexecutiontime.Thealgorithmthatcomputesf(p)basedonbi-sectionsearchisgiveninAlgorithm 2 Algorithm2Searchforf(p) INPUT:n,m,,p OUTPUT:framesizethatminimizesexecutiontimeundersamplingprobabilityp ifPemd(p,U)1do f2=df0+f1 2e; ifPemd(p,f2)
PAGE 51

Figure4-2. Framesizef(p)withrespecttosamplingprobabilitypwhenn=50,000,m=75,and=95%. Figure4-3. Executiontimef(p)tswithrespecttoenergycostnp. Thetradeoffbetweenthesetwoperformancemetricsiscontrolledbythesamplingprobabilityp.Ifwedecreasethevalueofp,wedecreasetheenergycost,butatthemeantimethevalueoff(p)mayhavetoincrease,whichincreasestheexecutiontime. 4.3.4MinimumEnergyCostWhenthesamplingprobabilitypistoosmall,thedetectionprobabilityPemd(p,f)willbesmallerthanforanyvalueoff.SuchasmallsamplingprobabilitycannotbeusedbyEMD.Wedesignabi-sectionsearchmethodinAlgorithm 3 tondthesmallestvalueofp,denotedaspopt,whichcansatisfyPemd(p,f)withaframesizenogreaterthan Figure4-4. Energy-timetradeoffcurveintherangep2[popt,pt]. 51

PAGE 52

theupperboundU.Thesamplingprobabilityreturnedbythealgorithmiswithinanerroroffromthetrueoptimalvaluepopt,whereisaparameterofthealgorithmthatcanbesetarbitrarilysmall.WhenEMDusespoptandf(popt),itsenergycostisminimized. Algorithm3Searchforpopt INPUT:n,m, OUTPUT:optimalsamplingprobabilitythatminimizesenergycost p0=0,p1=1,=0.01,U=1,000,000;nn400seconds whilep1)]TJ /F3 11.955 Tf 11.96 0 Td[(p0>do p2=dp0+p1 2e; ifPemd(p2,U)do p2=dp0+p1 2e; iff(p2)>f(p2+ 2)thenp0=p2elsep1=p2; endwhile returnf(p1)andp1; 52

PAGE 53

Figure4-5. Thevalueofptwithrespectton. Figure4-6. Thevalueofpoptwithrespectton. 4.3.6Energy-TimeTradeoff,TRP,andOfineComputationFig. 4-2 showstheenergy-timetradeoffcurve.Let'stakeacloserlookbyexaminingthesegmentofthecurvebetweenpoint(popt,f(popt))andpoint(pt,fopt)inthethirdplotofFig. 4-4 .Whenweincreasethevalueofpfrompopttopt,theenergycostoftheprotocolislinearlyincreased,whiletheexecutiontimeoftheprotocolisdecreased.Weshouldnotchoosep>ptbecausebothenergycostandexecutiontimewillincreasewhenthesamplingprobabilityisgreaterthanpt.TRP[ 46 ]isaspecialcaseofEMDwhenp=1.AsweseeinFig. 4-2 ,p=1isnotagoodchoiceforeitherenergyefciencyortimeefciency.Infact,itistheworstintermsofenergycost.Becausethecomputationofpopt,f(popt),pt,andfoptreliesonlyonthevaluesofn,mand,wecancalculatethemofineinadvance.Thevaluesofmandarepre-conguredaspartofthesystemrequirement.Hence,wecanpre-computepopt,f(popt),pt,andfoptinatableformatwithrespecttodifferentvaluesofn(forinstance,from100to100,000withanincrementstepof100),sothatthesevaluescanbelooked 53

PAGE 54

upduringonlineoperations.IttakestwohoursforustobuildsuchatableonaThinkpadT400laptop. 4.3.7FurtherEnergyReductionByselectingasamplingprobabilitybetweenptandpopt,EMDcangreatlyreducetheenergycostwhencomparingwithTRP.Wecanfurtherreducethetags'overallenergyconsumptionsuchthatanytagthatdoesnotrespondwillnotevenlistenforthepollingrequest.Weobservethatwhenwechangen,thevaluesofptandpoptremainlargelyconstants,asshowninFig. 4-5 andFig. 4-6 .Hence,theirvaluesaredeterminedbymand,whicharepre-determinedsystemparametersthatspecifythedetectionrequirement.Itmeansthatwecanprecomputethevaluesofptandpopt.Aslongasthedetectionrequirementspeciedbymanddoesnotchange,ptandpoptcanbeapproximatelyviewedasconstantseventhoughthenumberoftagsinthesystemchanges.Supposethedetectionrequirementmaybechangedonlyatthebeginningofeachday.Thereaderpicksasamplingprobabilityp,whichispt,poptoravaluebetweenthem.Itthendownloadsptoalltagsandsynchronizestheirclocks.Fortherestoftheday,thereaderdoesnothavetotransmitthesamplingprobabilityagain.Thetagswakeupatthetimeswhentheprotocolisscheduledforexecution.Eachtagmakesadecisionwithprobabilitypwhethertoparticipateinthepolling.Forthosethatdecidenotto,theygobacktothesleepmode.Theexpectednumberoftagsthatwillparticipateisnp.Thesetagsstayawaketoreceivethepollingrequest,andthenrespondinrandomly-selectedslotsinthetimeframeaftertherequest.Theactualimplementationofthesamplingprobabilitypisslightlydifferentfromwhat'sdescribedinSection 4.3.1 .Atthebeginningoftheday,thereaderdownloadsanintegerx=pXandarandomnumberrtoalltags,whereXisalargeconstant(say,216)andrisrelativelyprimetothesizeofthehash-bitring.Atthetimeoftheithprotocolexecutionduringtheday,thetagtakesahashvaluehfromthehash-bitring, 54

PAGE 55

startingattheithbitandtakingeveryrthbitdownthering,untillog2Xbitsaretaken.Itparticipatesinthepollingifh
PAGE 56

Theproblemisthatthemajorityofallslots,63.2%ormoreofthem,areeitheremptyslotsorcollisionslots.Theyaremostlywasted.Obviously,emptyslotsdonotcontributeanythinginmissing-tagdetection.Ifacollisionslotonlyhasmissingtags,detectionwillbesuccessfullymadebecausethereaderwillndthisexpectedbusyslottobeactuallyidle.However,whenthenumberofmissingtagsissmallwhencomparingwiththetotalnumberoftags,thechanceforacollisionslottohaveonlymissingtagsisalsosmall.Naturally,wewantaprotocoldesignthatensuresalargevalueof,muchlargerthan36.8%,becausemoresingletonslotsincreasedetectionpower.However,thevalueofinTRPisinfactmuchsmallerthan36.8%becauseTRPminimizesitsexecutiontimebyusingasfewtimeslotsaspossible,whichresultsinalargepercentageofcollisionslots.ThedetectionprobabilityofTRPisabout1)]TJ /F6 11.955 Tf 10.75 0 Td[((1)]TJ /F4 11.955 Tf 10.76 0 Td[()mbecauseeachofthemmissingtagshasaprobabilityoftomaptoasingletonslotandthusbedetected.1Asanexample,iftherequirementistodetectamissing-tageventwith99%probabilitywhen100tagsaremissing,TRPwillreduceitsframesizetosuchalevelthat=4.5%,justenoughtoensure99%detectionprobability.Thisleavesagreatroomforimprovement.Weshowthatanewprotocoldesign,differentfromthatofTRPandEMD,canreducetheframesizetoalevelthatismuchsmallerthantheycando,yetkeepatavaluemuchgreaterthan36.8%.Ourdesign,calledMultiple-SeedMissing-tagDetectionprotocol(MSMD),turnsmostempty/collisionslotsintosingletons.Thereisacompoundeffectofsuchanewdesignwhenitiscoupledwithsampling:Suppose=99%andm=100,sameasinthepreviousparagraph.Undersampling,thedetectionprobabilityis1)]TJ /F6 11.955 Tf 12.26 0 Td[((1)]TJ /F3 11.955 Tf 12.26 0 Td[(p)mbecauseeachofthemmissingtagshasaprobabilityofptobesampledandmappedtoasingletonslot. 1Toquicklygettothepointwithoutdealingwithtoomuchdetail,weignorethesmallcontributionofcollisionslotsindetection. 56

PAGE 57

Ifourprotocoldesigncanimproveto90%,wewillbeabletosetp=5%.Withsuchasamplingprobability,weachievemuchbetterenergyefciencybecauseonly5%ofalltagsparticipateineachprotocolexecution.Wealsoachievefarbettertimeefciencybecause,withmuchfewertagstransmitting,thechanceofcollisionisreducedandafewernumberoftimeslotsareneededtoensureacertainlevelofsingletons. 4.4.2HashFunctionThereexistmanyefcienthashfunctionsintheliterature.Inordertokeepthetagcircuitsimple,webuildourhashfunctionontopofthesimpleschemein[ 24 ]usingaringofpre-storedrandombits:Beforeatagisdeployed,anofinerandomnumbergeneratorusestheIDofatagasseedtoproduceastringofpseudo-randombits,whicharestoredinthetag.Thebitsformalogicalring.Afterdeployment,thetaggeneratesahashvalueH(id,s)byreturningacertainnumberofbitsafterthesthbitinthering,whereidisthetagIDandsisagivenhashseedthatcanalterthehashoutput.ThishashoutputispredictablebyaRFIDreaderthatknowsthetagIDandtheseeds.Moresophisticatedhashimplementationscanbedesignedbasedonaringofpseudo-randombits.Forexample,wemayinterpretsasaconcatenationofaagxandtworandomnumbers,r1andr2.Toproduceahashoutput,wegoclockwisealongtheringifx=0orcounterclockwiseifx=1.Wethenoutputther1thbitonthering,andthenoutputonemorebitaftereveryr2bitsonthering.Ifthehashoutputisrequiredtobeinarange[0,y),wersttakeasufcientnumberofhashbitsasdescribedaboveandthenperformmoduloy.Thishashfunctioniseasytoimplementinhardwareandthussuitablefortags.Butitcanonlyproducealimitednumberofdifferenthashvalues,dependingonthesizeofthering.Itisnotsuitableforaprotocolwhoseoperationsrequireeachtagtoproducealargenumberofdifferenthashvalues,butitworkswellforaprotocolthatonlyrequireseachtagtoproduceafewindependenthashvalues. 57

PAGE 58

4.4.3BasicIdeaWehaveknownthatunderarandommappingfromtagstoslots,anarbitraryslotonlyhasaprobabilityofupto36.8%tobeasingleton.Now,ifweseparatelyapplytwoindependentrandommappingsfromtagstoslots,aslotwillhaveaprobabilityofupto1)]TJ /F6 11.955 Tf 12.01 0 Td[((1)]TJ /F6 11.955 Tf 12 0 Td[(36.8%)260.1%tobeasingletoninoneofthetwomappings.Ifweseparatelyapplykindependentmappingsfromtagstoslots,ithasaprobabilityof1)]TJ /F6 11.955 Tf 11.99 0 Td[((1)]TJ /F6 11.955 Tf 11.99 0 Td[(36.8%)ktobeasingletoninoneofthekmappings.Thevalueof1)]TJ /F6 11.955 Tf 12.99 0 Td[((1)]TJ /F6 11.955 Tf 12.99 0 Td[(36.8%)kquicklyapproachesto100%asweincreasek.Itiseasytogeneratemultiplemappings.Inthedetectionrequest,theRFIDreadercanbroadcastkseeds,s1,s2,...,sk,totags.Eachseedsicorrespondstoadifferentmapping,whereatagismappedtoaslotindexedbyH(id,si),whichisahashfunctionsuchas[ 24 ]thattakesanIDandaseedtoproduceanoutput(belongingtoarequiredrangethroughmodulooperation).Aslotmaybeasingletonunderonemapping,butacollisionslotunderothermappings.Differentslotsmaybesingletonsunderdifferentmappings.Tomaximizethenumberofsingletons,thereaderwiththeknowledgeofalltagIDsandallseedsselectsamapping(i.e.,aseed)foreachslot,suchthattheslotcanbeasingleton.Thereaderalsomakessurethateachtagisassignedtoasingletononlyonce.Fromeachslot'spointofview,aspecicseedisusedtomaptagstoit.Fromthewholesystem'spointofview,multipleseedsareusedtomapdifferenttagstodifferentslots.Inourprotocol,thereaderdeterminessystemparameters,includingthesamplingprobabilitypandtheframesizef.Afterselectingkrandomseeds,thereaderchoosesaseedforeachslotandconstructsaseed-selectionvectorV(orselectionvectorforshort),whichcontainsfselectors,oneforeachslotinthetimeframe.Eachselectorzhasarangeof[0,k].Ifz>0,itmeansthatthezthseed,i.e.,sz,shouldbeusedforitscorrespondingslot.Ifz=0,itmeansthattheslotisnotasingletonunderanyseed. 58

PAGE 59

Figure4-7. InPhasetwo,thereaderbroadcaststheseed-selectionsegments,V1throughVf=l,oneatatime.EachsegmentViisimmediatelyfollowedbyasub-frameFioflslots,duringwhichthetagstransmit. Finally,thereaderbroadcaststheselectionvectortothetags.Basedontheselectors,eachtagdetermineswhichslotitshouldusetorespond.Wewilladdresstheproblemsofhowtochoosetheoptimalsystemparameters,pandf,andhowthenumberkofseedswillaffecttheprotocolperformanceinSection 4.5 .Beforewedescribetheoperationsoftheprotocol,weintroducetheconceptofsegmentation.Inourdesign,theaboveideaisactuallyappliedsegmentbysegment. 4.4.4SegmentationTheseed-selectionvectorhasfselectors,eachofwhicharedlog2(k+1)ebitslong.fmaybetoolargeforthewholevectortotinasingleslot.Forexample,ifk=7,eachselectoris3bitslong.IfweusethesameslotTtagforcarryinga96-bitIDtocarrytheselectionvector,itcanonlyaccommodate32selectors.Whenfismorethanthat,wehavetodividetheselectionvectorinto96-bitsegments,sothattheycantinTtagslots.Eachsegmentcontainsl=96 dlog2(k+1)eselectors.Thetotalnumberofseed-selectionsegmentsaref l,andthejthsegmentisdenotedasVj.Sincewedividetheselectionvectorintosegments,wealsodividethetimeframeintosub-frames,eachcontaininglslotsaccordingly.Thejthtimesub-frameisdenotedasFj.Thisallowsourprotocoltodealwithonesub-frameatatime. 4.4.5ProtocolOverviewOurprotocolconsistsoftwophases.Phaseoneperformstagassignment,wherethereaderidentiesthesetofsampledtags,andassignsthesampletagstothesub-framesuniformlyatrandom.Thesubsetofsampledtagsthatareassignedtothe 59

PAGE 60

Figure4-8. Arrowsrepresentthemappingfromtagstoslotsbasedonhashfunctions.Amongthem,thickarrowsrepresenttheassignmentoftagstoslots.Inthisexample,k=2. jthsub-frameisdenotedasNj.Foreachsub-frameFj,thereaderselectsaseedforeachofitsslots,constructstheseed-selectsegmentVj,andmapsthetagsinNjtoslotsinFjusingtheselectedseeds.Phasetwoperformsmissing-tagdetection.Thereaderbroadcaststheseed-selectionsegmentsoneafteranother,eachinaslotofTtag.Eachseed-selectionsegmentisfollowedbyatimesub-frameoflslots,eachofwhichisTshortlong.ThetagsinNjwillrespondintheseslots.Eachtagonlyneedstobeactiveduringitssub-frame,whichconservesenergy.TheexchangebetweenthereaderandtagsinPhasetwoisillustratedinFigure 4-7 4.4.6PhaseOne:TagAssignmentPhaseoneconsistsofthreesteps,whichareexplainedbelow.AnillustrativeexamplecanbefoundinFig. 4-8 60

PAGE 61

4.4.6.1DeterminingsampledtagsThereaderstartsPhaseonebyuniquelyidentifyingthesetofparticipatingtagsthroughsampling.Toimplementthesamplingprobabilityp,thereaderbroadcastsanintegerx=dpXeandaprimenumberq,whereXisalarge,pre-conguredconstant(e.g.,216).Duringtheithroundofprotocolexecution,atagissampledifandonlyifthehashresultH(id,qi),whichisapseudo-randomnumberintherangeof[0,X),issmallerthanx,whereidisthetag'sID.Afterreceivingxandq,eachtagcanpredictwhichroundsofprotocolexecutionitwillparticipate.Sincetheprotocolisscheduledtoexecuteperiodicallywithpre-denedintervals,eachtagknowswhenitshouldwaketoparticipate.Thereader,withtheknowledgeofalltaginformation,canpredictwhichtagsaresampledforeachprotocolexecution. 4.4.6.2Assigningsampledtagstosub-framesWhenassigningsampledtagstotimesub-frames,thereaderselectsanadditionalrandomseeds,whichisdifferentfroms1,...,sk.Foreachsampledtag,thereaderproducesahashoutputH(id,s)andassignsthetagtotheH(id,s)thsub-frame,whereidisthetag'sIDandtherangeofH(id,s)is[0,f l).Notethateachtagwillknowwhichsub-frameitisassignedto,afteritreceivessinthedetectionrequestbroadcastbythereaderatthebeginningofPhasetwo. 4.4.6.3Determiningseed-selectionsegmentsEachseed-selectionsegmentisdeterminedindependently.AllselectorsinVjareinitializedtozerosasshowninFig. 4-8 (b).Thereaderbeginsbyusingtherstseeds1tomaptagsinNjtoslotsinFj,asshowninFig. 4-8 (c).ForeachtaginNj,thereaderproducesahashoutputH(id,s1)andmapsthetagtotheH(id,s1)thslotinFj,whereidisthetag'sIDandtherangeofH(id,s1)is[0,l).Aftermapping,thereaderndssingletonslots.Eachsingletonhasoneandonlyonetagmappedtoitasanexample,therstandthirdslotsinFig. 4-8 (c).Weassignthetagtotheslotsothatitwilltransmit 61

PAGE 62

intheslot,freeofcollision,duringPhasetwo.ThereadersetsthecorrespondingselectorinVjtobe1,meaningthattherstseeds1shouldbeusedforthisslot.Theslotisnowcalledausedslot,andthesoletagmappedtoitwillbecalledanassignedtag.Thereaderrepeatstheaboveprocesswithotherseeds,oneatatime,fortheremainingmappings.Foreachmapping,weonlyconsidertheslotswhoseselectorshavenotbeendeterminedyetandonlyconsiderthetagsthathavenotbeenassignedtoanyslotsyet,asshownbyFig. 4-8 (d).Inotherwords,theusedslotsandtheassignedtagswillnotbeconsidered.Forasingletonslotthatisfoundusingseedsi,thecorrespondingselectorinVjwillbesettobei.Afterallkmappings,ifthevalueofaselectorinVjremainszero,itmeansthatthecorrespondingslotinFjisnotasingletonunderanyseed.Asanalattempttoutilizetheseunusedslots,ifthereexistunassignedtagsinNj,thereaderrandomlyassignstheunassignedtagstounusedslots.Morespecically,itchoosesanadditionalrandomseeds0andproducesahashoutputH(id,s0)toassigneachtagthatisnotassignedyettotheH(id,s0)thunusedslot,whereidisthetag'sID.Incasethatonlyonetagisassignedtoanunusedslot,wewillhaveanextrasingleton,asshowninFig. 4-8 (e).Sincethewholetag-to-slotassignmentispseudo-random,thereaderknowswhichunusedslotswillbecomesingletons.AswewillseelaterinPhasetwo,afterreceivings1,...,sk,eachtagwillknowwhetheritisassignedtoaslot.Ifnot,fromthereceiveds0,itwillknowwhichunusedslotitisassignedto. 4.4.7PhaseTwo:Missing-TagDetectionAtthebeginningofthisphase,thereaderbroadcastsadetectionrequest,whichisfollowedbyatimeframeforsampledtagstorespond.Thedetectionrequestconsistsofaframesizefandasequenceofseeds,s,s1,...,sk,ands0.Thetimeframeisdividedintosub-frames.Beforeeachsub-frameFj,thereaderbroadcaststhecorrespondingseed-selectionsegmentVjinasingletag-IDslotTtag.Itisfollowedbylshortslots(Tshort)ofthesub-frame,duringwhichthetagsinNjcanrespond.Recallthateach 62

PAGE 63

selectionsegmentis96bitslong.Ifk=7,asegmenthasl=96 log2(7+1)=32selectors,andthuseachtimesub-framehas32slots.Consideranarbitrarytagt.Itwakesuptoparticipateinascheduledprotocolexecutionthatitissampledfor.Aftertreceivesthedetectionrequestfromthereader,itusesH(id,s)todeterminewhichsub-frameitisassignedto.Withoutlossofgenerality,letthesub-framebeFj.ThetagsetstimertowakeupbeforeFjbegins.Afterreceivingtheseed-selectionsegmentVj,tagtusesH(id,s1)tondoutwhichtimeslotitismappedbyseeds1.ItthencheckswhetherthecorrespondingselectorinVjis1.Iftheselectoris1,accordingtotheconstructionofVjinSection 4.4.6.3 ,tmustbethesoletagthatismapped(andassigned)tothisslotunders1.Iftheselectorisnot1,itmeansthats1shouldnotbeusedtomapanytagtothisslot.Inthelattercase,twillmoveontootherseedsandrepeatthesameprocesstodetermineifitisassignedtoaslot.Ifso,itwilltransmitduringthatslot.Otherwise,iftisnotassignedtoaslotafterallkseeds,itwillmakeanalattemptbyndingoutallunusedslots(whosecorrespondingselectorsinVjarezeros)andusingH(id,s0)asindextoidentifyanunusedslottotransmit.Insummary,afterPhaseone,thereaderknowswhichsub-frameeachsampledtagisassignedto,whichsloteachsampledtagisexpectedtotransmit,whichslotsareexpectedtobesingletons,andwhichslotsareexpectedtobecollisionslots(duetothenalattemptusings0).AfterPhasetwo,ifanexpectedsingleton/collisionslotturnsouttobeidle,thereaderdetectsamissing-tagevent.Becausemultiplemappingsreducethenumberofempty/collisionslots,bothenergyefciencyandtimeefciencyaregreatlyimproved,aswewilldemonstrateanalyticallyandbysimulationsinthefollowingsections. 4.5Energy-TimeTradeoffinProtocolCongurationWestudytheenergy-timetradeoffofourprotocolandshowhowtocomputethesystemparameters. 63

PAGE 64

4.5.1ExecutionTimeandEnergyCostTheprotocolexecutiontimeincludesthetimeforthereadertotransmitadetectionrequest,thetimeforthereadertotransmittheseed-selectionvectoroffselectors,andthetimeframeoffslotsfortagstotransmit,wheretheseed-selectionvectorisdividedintosegmentsandthetimeframeisdividedintosubframes.Therequestonlycarriesafewparameters.Itstimeisnegligiblewhencomparingwiththetimeframeandtheseed-selectionvectoriffislarge.Hence,theprotocolexecutiontimeisroughlyproportionaltof.Toinvestigatetheenergy-timetradeoffinrelativeterms,wecharacterizetheprotocolexecutiontimebyusingtheframesizef.Asmallervalueoffmeansashorterprotocolexecutiontime.Theactualexecutiontime,measuredinseconds,willbestudiedthroughsimulationsinSection 4.7 .Thecomputationateachtagismostlyhashing.Thehashfunctioncanbemadeverysimple,suchas[ 24 ]wherehashoutputisproducedbyselectingacertainnumberofbitsfromapre-storedbitring.Moreover,oncethetagsreceivethehashseedsfromthereader'sdetectionrequest,theycanproducetheneededhashvaluesaheadoftime,andalltagsdosoinparallel,whilethereaderissendingitsseed-selectionvector.Ourprotocoldesignpushesmostofitscomplexitytothereader.Thetags'operationissimple:Atagwakesuptoparticipateinascheduledprotocolexecutionthatitissampledfor.Itreceivesthedetectionrequest,determineswhichsub-frameitisassignedto,wakesupagainbeforethesub-framestarts,receivesthe96-bitseed-selectionsegment,determineswhichslotitisassignedto,andtransmitasignalinthatslot.Becauseeachparticipatingtagperformssimilaroperation,theenergycosttoeachparticipatingtagisalsosimilar.Thetotalenergycostamongalltagsforeachprotocolexecutionisproportionaltothenumberofparticipatingtags;notethatdifferenttagswillbesampledtoparticipateindifferentprotocolexecutionsuniformlyatrandom.Hence,wemaycharacterizetheenergycostofaprotocolexecutionbyusingtheexpected 64

PAGE 65

numberofparticipatingtags,pn,whichisinturnproportionaltothesamplingprobabilityp. 4.5.2DetectionProbabilityTondthedetectionprobabilityafteroneprotocolexecution,weneedtorstderivetheprobabilityforanarbitrarysampledtagttobeassignedtoasingletonslotduringPhaseone.Therearekmappings.LetPibetheprobabilitythattagtisassignedtoasingletonslotaftertherstimappings.Letnbethetotalnumberoftagsandn0bethenumberofsampledtagsthataremappedtothesamesub-frameastdoes.Assumethehashfunctionassignssampledtagstosub-framesuniformlyatrandom.n0followsabinomialdistribution,Bino(n,pl f),i.e., Probfn0=jg=nj(pl f)j(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pl f)n)]TJ /F7 7.97 Tf 6.59 0 Td[(j.(4)P0=0.WederivearecursiveformulaforPi,1ik.Afterthersti)]TJ /F6 11.955 Tf 9.88 0 Td[(1mappings,therearetwocases.Case1:tagthasbeenassignedtoaslot;theprobabilityforthistohappenisPi)]TJ /F5 7.97 Tf 6.59 0 Td[(1.Case2:tagthasnotbeenassignedtoaslot;theprobabilityforthiscaseis1)]TJ /F3 11.955 Tf 11.96 0 Td[(Pi)]TJ /F5 7.97 Tf 6.59 0 Td[(1.Wefocusonthesecondcasebelow.Intheithmapping,theslotthattagtismappedtohasaprobabilityof(1)]TJ /F7 7.97 Tf 13.2 5.94 Td[(n0Pi)]TJ /F21 5.978 Tf 5.76 0 Td[(1 l)tobeunused.Eachoftheothern0)]TJ /F6 11.955 Tf 11.84 0 Td[(1tagshasaprobability(1)]TJ /F3 11.955 Tf 11.84 0 Td[(Pi)]TJ /F5 7.97 Tf 6.58 0 Td[(1)tobeunassigned.Ifitisunassigned,thetaghasaprobabilityof1 ltobemappedtothesameslotastdoes.Hence,theprobabilityp0fortagttobetheonlyonethatismappedtoanunusedslotis p0=(1)]TJ /F6 11.955 Tf 11.95 0 Td[((1)]TJ /F3 11.955 Tf 11.95 0 Td[(Pi)]TJ /F5 7.97 Tf 6.59 0 Td[(1)1 l)n0)]TJ /F5 7.97 Tf 6.59 0 Td[(1(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(n0Pi)]TJ /F5 7.97 Tf 6.58 0 Td[(1 l).(4) 65

PAGE 66

RecallthatweareconsideringCase2here.Combiningbothcases,wehave Pi=Pi)]TJ /F5 7.97 Tf 6.59 0 Td[(1+(1)]TJ /F3 11.955 Tf 11.95 0 Td[(Pi)]TJ /F5 7.97 Tf 6.59 0 Td[(1)nXj=1Probfn0=jgp0=Pi)]TJ /F5 7.97 Tf 6.59 0 Td[(1+(1)]TJ /F3 11.955 Tf 11.95 0 Td[(Pi)]TJ /F5 7.97 Tf 6.59 0 Td[(1)nXj=1nj(pl f)j(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pl f)n)]TJ /F7 7.97 Tf 6.59 0 Td[(j(1)]TJ /F6 11.955 Tf 11.96 0 Td[((1)]TJ /F3 11.955 Tf 11.96 0 Td[(Pi)]TJ /F5 7.97 Tf 6.58 0 Td[(1)1 l)j)]TJ /F5 7.97 Tf 6.59 0 Td[(1(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(jPi)]TJ /F5 7.97 Tf 6.59 0 Td[(1 l),(4)wheretherstitemontherightsideistheprobabilityforatagtobeassignedtoaslotbythersti)]TJ /F6 11.955 Tf 12.5 0 Td[(1mappingsandtheseconditemistheprobabilityforthetagtobeassignedtoaslotbytheithmapping.TheprobabilityfortagttobeassignedtoaslotafterallkmappingsisPk.Afterthekmapping,wehaveanalattempt,inwhichanunassignedtagmaybemappedtoasingletonslotoracollisionslot.Ifthetagismappedtoacollisionslot,itishighlyunlikelythatalltagsinthatslotwillbemissingbecausetheparametermistypicallysetfarsmallerthann.Hence,thecontributionofcollisionslotstomissing-tagdetectioncanbeignored.Whenthetagismappedtoasingletonslot,detectionwillbemadeifthetagismissing.Therefore,thenalmappinghasnodifferencefromthepreviousmappings.TheprobabilityfortagttotransmitinasingletonslotisPk+1,whichcanbecomputedrecursivelyfrom( 4 ).Eachofthemmissingtagshasaprobabilityptobesampled.Whenthetagissampled,ithasaprobabilityofPk+1tobeassignedasingletonslot.Whenthathappens,sinceamissingtagcannottransmit,thereaderwillobserveanidleslotinstead,resultinginthedetection.Therefore,thedetectionprobabilityofMSMD,denotedasPmsmd(p,f),isPmsmd(p,f)=1)]TJ /F6 11.955 Tf 11.96 0 Td[((1)]TJ /F3 11.955 Tf 11.96 0 Td[(pPk+1)m. (4)ThevalueofPmsmd(p,f)notonlydependsonthechoiceofpandf,butalsodependsonn,mandk,whicharenotincludedinthenotationforsimplicity.Thevaluesofpand 66

PAGE 67

Figure4-9. DetectionprobabilityPmsmd(p,f)withrespecttotheframesizefwhenn=50,000,m=100,k=3,andp=5%. faredeterminedbythereaderandbroadcasttotags.Theycontroltheenergy-timetradeoffaswewillrevealshortly.Thevaluesofn,mandkarepre-known,wherenisknownbecauseitissimplythenumberoftagsthatthereaderexpectstobeinthesystem,misknownasagivenparameterinthedetectionrequirement,andkisdeterminedbeforethetagsaredeployed.EMD[ 29 ]isaspecialcaseofMSMDwithk=1andwithoutthenalattempt.Hence,thedetectionprobabilityofEMD,denotedasPemd(p,f),isPemd(p,f)=1)]TJ /F6 11.955 Tf 11.95 0 Td[((1)]TJ /F3 11.955 Tf 11.96 0 Td[(pP1)m. (4)TRP[ 46 ]isaspecialcaseofEMDwithp=1.Namely,samplingisturnedoff. 4.5.3Energy-TimeTradeoffCurveWecannotarbitrarilypicksmallvaluesforpandf.TheymustsatisfytherequirementPmsmd(p,f).Subjecttothisconstraint,weshowthatthevaluesofpandfcannotbeminimizedsimultaneously.Thechoiceofpandfrepresentsanenergy-timetradeoff.Ifwexthevalueofp,Pmsmd(p,f)becomesafunctionoff.ThesolidlineinFig. 4-9 showsanexampleofthecurvePmsmd(p,f)withrespecttofwhenn=50,000,m=100,k=3,andp=5%.BecausePmsmd(p,f)isanincreasingfunction,theminimumvalueoffthatsatisestherequirementPmsmd(p,f)canbefoundbysolvingthefollowingequation,Pmsmd(p,f)=. 67

PAGE 68

Figure4-10. Energy-timetradeoffcurve,i.e.,framesizef(p)withrespecttosamplingprobabilityp,whenn=50,000,m=75,k=3,and=95%. Thesolutionisdenotedasf.SeeFig. 4-9 forillustration.Foreachdifferentsamplingprobabilityp,wecancomputethesmallestusableframesizefthatsatisesPmsmd(p,f).Hence,fcanbeconsideredasafunctionofp,denotedasf(p).ApracticalRFIDsystemmayconsideraframesizebeyondacertainupperboundUtobeunacceptableduetoexcessivelylongexecutiontime.Inaddition,fmustbeaninteger.Consideringthesefactors,wegiveamoreaccuratedenitionoffbelow.f(p)=minffjPmsmd(p,f)^fU,f2I+g. (4)Wecanndthevalueoff(p)throughbi-sectionsearch.TheleftplotinFig. 4-10 showsthecurveoff(p)whenn=50,000,m=75,k=3,and=95%.Wecallittheenergy-timetradeoffcurve.Eachpoint,(p,f(p)),representsanoperatingpointwhoseenergycostismeasuredasnpparticipatingtagsandwhosetimeframeconsistsoff(p)slots.Thesymbolsintheplotwillbeexplainedlater.Theenergy-timetradeoffiscontrolledbythesamplingprobabilityp.Ifwedecreasethevalueofp,wedecreasetheenergycost,butatthemeantimethevalueoff(p)mayhavetoincrease,whichincreasestheexecutiontime. 4.5.4MinimumEnergyCostWhenthesamplingprobabilitypistoosmall,thedetectionprobabilityPmsmd(p,f)willbesmallerthanforanyvalueoff.Suchasmallsamplingprobabilitycannotbe 68

PAGE 69

Figure4-11. Energy-timetradeoffcurveintherangep2[popt,pt],whichcorrespondstothecurvesegmenttotheleftofthedashedlineinFig. 4-10 Figure4-12. Thevalueofptwithrespectton. used.Wecanusebi-sectionsearchtondthesmallestvalueofp,denotedaspopt,whichcansatisfyPmsmd(p,f)withaframesizenogreaterthantheupperboundU.Whenpoptandf(popt)areused,theenergycostisminimized. 4.5.5MinimumExecutionTimeFromtheenergy-timetradeoffcurve(theleftplotinFig. 4-10 ),wecanndthesmallestvalueoff(p),denotedasfopt,thatminimizestheexecutiontime.fopt=minff(p)jpoptp1g (4) Figure4-13. Thevalueofpoptwithrespectton. 69

PAGE 70

Letptbethecorrespondingsamplingprobability.Namely,Pmsmd(pt,fopt)=.Thevaluesoffoptandptcanbedeterminedthroughbi-sectionsearch.Whenptandfoptareused,theprotocolexecutiontimeisminimized.Weamplifythesegmentoftheenergy-timetradeoffcurvebetweenpoint(popt,f(popt))andpoint(pt,fopt)intherightplotofFig. 4-10 .Whenweincreasethevalueofpfrompopttopt,theenergycostoftheprotocolislinearlyincreased,whiletheexecutiontimeoftheprotocolisdecreased.Weshouldnotchoosep>ptbecausebothenergycostandexecutiontimewillincreasewhenthesamplingprobabilityisgreaterthanpt. 4.5.6OfineComputationBecausethecomputationofpopt,f(popt),pt,andfoptreliesonlyonthevaluesofn,m,andk,wecancalculatethemofineinadvance.Thevaluesofmandarepre-conguredaspartofthesystemrequirement.Thevalueofkisdeterminedbeforetagdeployment.Hence,wecanpre-computepopt,f(popt),pt,andfoptinatableformatwithrespecttodifferentvaluesofn,sothatthesevaluescanbelookedupduringonlineoperations.Whenperformingsuchcomputation,weobservethatwhenwechangen,thevaluesofptandpoptremainlargelyconstants,asshowninFig. 4-12 andFig. 4-13 .Hence,theirvaluesareactuallydeterminedbym,andk.Itmeansthataslongasthedetectionrequirementspeciedbymanddoesnotchange,ptandpoptcanbeapproximatelyviewedasconstantseventhoughthenumberoftagsinthesystemchanges.Supposethevaluesofmandmaybechangedonlyatthebeginningofeachhour.Thereaderpicksasamplingprobabilityp,whichispt,poptoravaluebetweenthem.Itthendownloadsptoalltagsandsynchronizestheirclocks.Fortherestofthehour,thereaderdoesnothavetotransmitthesamplingprobabilityagain. 4.5.7ConstrainedLeast-Time(orLeast-Energy)ProblemTheenergy-constrainedleast-timeproblemistominimizetheprotocol'sexecutiontime,subjecttoadetectionrequirementspeciedbymandandanenergyconstraint 70

PAGE 71

speciedbyanupperbounduontheexpectednumberoftagsthatparticipateineachprotocolexecution.Tominimizeexecutiontime,weneedtoreducetheframesizeasmuchaspossible.Ourpreviousanalysishasalreadygiventhesolutiontothisproblem,whichissimplyf(u n),whereu nisthemaximumsamplingprobabilitythatwecanuseundertheenergyconstraint.Thetime-constrainedleast-energyproblemistominimizethenumberoftagsthatparticipateinprotocolexecution,subjecttoadetectionrequirementspeciedbymandandanexecutiontimeconstraintspeciedbyanupperboundu0ontheframesize.Asolutioncanbedesignedbyfollowingasimilarprocessaswederivef(p)inSection 4.5.3 :Startingfrom( 4 ),ifwexf=u0,Pmsmd(p,f)becomesafunctionofp.Wecanusebi-sectionsearchtondpthatmeetsPmsmd(p,f)=^ppt. 4.5.8ImpactofkWestudyhowthenumberkofhashseedswillaffecttheprotocol'sperformance.Figure 4-14 comparestheenergy-timetradeoffcurvesofEMDandMSMDwithk=3,7,15,respectively.RecallthatEMDisaspecialcaseofMSMDwithonehashseedandTRPisaspecialcaseofEMDwithp=1,representedbyapointonthecurveofEMDasshowninthegure.ForMSMD,whenk=3,eachseedselectorneeds2bits;recallthatthevaluezeroisreservedfornon-singletonslots.Whenk=7,eachselectorneeds3bits.Whenk=15,eachselectorneeds4bits.2InFigure 4-14 ,alowercurveindicatesbetterperformancebecause,foranysamplingprobability,itsframesizeissmaller,i.e.,itsexecutiontimeissmaller.Alternatively,itcanbeinterpretedas,foranyframesize,itssamplingprobabilityissmaller,i.e.,itneedsfewertagstoparticipateineachprotocolexecution.Clearly,MSMD 2Onemayaskwhywedonotusek=8orothervalues.Thereasonisthateachselectorneeds4bitsevenwhenk=8.Inthatcase,weshouldcertainlychoosek=15forbetterperformance. 71

PAGE 72

Figure4-14. Energy-timetradeoffcurvesofEMDandMSMDunderdifferentkvalues,whenn=50,000,m=100,andp=5%. signicantlyoutperformsEMDandTRP.Askincreases,theperformanceofMSMDimproves.However,theamountofimprovementshrinksrapidly,demonstratedbythesmallgapbetweenk=7andk=15.Whenwefurtherincreasekto31using5-bitselectors,theimprovementbecomesnegligible.Increasingthevalueofkdoesnotcomeforfree;largerselectorsmeanmoreoverhead.Forone,ittakesmoretimeforthereadertobroadcasttheseed-selectionvector.Therefore,webelievek=7isagoodchoiceinpracticebecausetheperformancegainbeyondthatisverylimited. 4.6MSMDoverUnreliableChannelsWenowconsiderunreliablechannels.TheintuitionbehindEMDandMSMDisthataslotwillbefoundidleifthetagstransmittingintheslotareallmissing.Itistrueifnoerroroccursinthisslot.Inrealitythough,thecommunicationbetweenareaderandtagsis,tovaryingdegrees,subjecttonoise/interferenceintheenvironment,whichmaycorruptslots,forexample,turningawould-beemptyslottoabusyslot.Table 4-2 givesdifferentpossibilitiesofcorruptedslotsandtheirconsequences.Intherstrowofthistable,weassumethatatagtismissing.Then,theslotthattismappedtoshouldbecomeidleduringtheprotocolexecution.However,thisslotmaybecorruptedandturnouttobeabusyslot.Inthiscase,evenifatagthatissupposedtotransmitinaslotismissing,thereadercanstillsensearesponseinducedbychannelerror,resultinginundetectionofamissingtag.Ifallslotsassignedtomissingtagshappentobecorrupted,thereaderwillfailtodetectthemissing-tagevent.Thesecondrow 72

PAGE 73

Table4-2. Theimpactofcorructedslots OriginalslotAftercorruptionTagstatusReaderthinks idlebusymissingnotmissingbusybusynotmissingnotmissing illustratestheimpactofchannelerrorwhentispresentinthesystem.Recallthatweonlydistinguishtwostatesforaslot:idle(nosignaldetectedinthechannel)orbusy(signaldetected).Iftisnotmissing,theoriginalslotshouldbebusy.Sincenoiseorinterferenceintheenvironmentmaychangethesignalbutisunlikelytocancelthesignaloutaltogether,theslotshouldremainabusyslot.Inthiscase,noise/interferedoesnotcauseharm.Weevaluatetheimpactofchannelerrorundertwodifferentmodels:randomerrormodelandbursterrormodel.Theformercanbecharacterizedbyaparametererr,whichdenotestheprobabilityforeachslottoincurerror.However,itmayhappenthatchannelintroduceserrorinburstsforconsecutiveslotsratherthanatrandom.Forexample,communicationsbetweenareaderandtagsmaybeinterferedbyelectromagneticemissionfromnearbydevicesthatsharethesamefrequencyband.Whenthosedevicesaretransmitting(e.g.,sendingapacket),theirsignalskeepthechannelbusyforasmallperiodoftimeuntiltheystop.Astheinterferingtransmissionsareturnedonandoff,itcausesburstsoferrortotheRFIDsystem,whichischaracterizedbythebursterrormodel.Westressthattheerrormodelsareappliedonlytothetag-to-readerlinkinordertosimplifytheanalysis.Erroronthereader-to-taglinkisaddressedseparatelyasfollows:ThetransmissionfromthereadercarriesaCRCchecksum.Forexample,the96-bitseed-selectionsegmentmaycontaina16-bitCRCchecksumandusetheremaining80bitstoencodeseedselectors.Whenatagreceivesthetransmissionfromthereader,itcomputesaCRCbasedonthereceivedinformationandthencomparestheresultwiththereceivedCRC.Iftheyarethesame,thetagperformstheoperationaccordingtothe 73

PAGE 74

protocol.Otherwise,thetagwillnotparticipatefurtherintheprotocolexecution,andinsteaditwilltransmititsIDtothereadertoannounceitspresenceaftertheprotocolexecution.Thisaddsadditionalexecutiontimeandenergycost.Butiftheerrorratioissmall,theadditionaloverheadwillbesmall. 4.6.1MSMDunderRandomErrorModelIntherandomerrormodel,theimpactofchannelerrorischaracterizedbyaparametererr,whichistheprobabilityforaslottoincurerror.Forexample,iferr=5%,awould-beidleslothasachanceof5%tobecomebusy.MSMDunderthismodeliscalledMSMD-re.FromSection 4.5.2 ,weknowthateachofthemmissingtagshasaprobabilityofpPk+1tobedetectedinMSMD.Then,theprobabilityforamissingtagtobedetectedundertherandomerrormodelispPk+1(1)]TJ /F3 11.955 Tf -446.09 -23.91 Td[(err).Therefore,thedetectionprobabilityofMSMD-re,denotedasPmsmd)]TJ /F7 7.97 Tf 6.58 0 Td[(re(p,f,err),isPmsmd)]TJ /F7 7.97 Tf 6.59 0 Td[(re(p,f,err)=1)]TJ /F6 11.955 Tf 11.96 0 Td[((1)]TJ /F3 11.955 Tf 11.96 0 Td[(pPk+1(1)]TJ /F3 11.955 Tf 11.96 0 Td[(err))m. (4)Clearly,MSMDisaspecialcaseofMSMD-rewitherr=0.ForMSMD-re,thecomputationofpopt,f(popt),pt,andfoptissimilartothatofMSMDexceptthat( 4 )isreplacedwith( 4 ).MSMD-reusesthesameofinecomputationprocessasdescribedinSection 4.5.6 andfollowsthesameprotocol,onlywithmodiedparametersthatconsidertheimpactofchannelerror. 4.6.2MSMDunderBurstErrorModelWenowconsiderthebursterrormodel.Accordingto[ 13 ],thenumberofburstscanbeapproximatedasPoissondistribution.Wegivebriefdescriptionbelowandreadersarereferredto[ 13 ]fordetails.Theprobabilitydensityfunctionforthenumberofburstsisgivenbyh(x)=1Xi=0i i!e)]TJ /F9 7.97 Tf 6.59 0 Td[((x)]TJ /F3 11.955 Tf 11.95 0 Td[(i), (4)whereistheaveragenumberofbursts,and()istheDiracDeltaFunction[ 1 ]. 74

PAGE 75

Accordingtoconvolutionalcodesandtrelliscodemodulations,theprobabilitydensityfunctionforthenumberoferrorsinaburstcanbederivedbyErlangdensityofsecondorder.TheErlangdistributionofsecondorderisgc(z)=(2)2ze)]TJ /F5 7.97 Tf 6.59 0 Td[(2z, (4)whereistherateparameter.Becausetherandomvariableforthenumberoferrorsinaburstassumesonlydiscretevalues[ 13 ],theprobabilitydensityfunctioncanbeobtainedbydiscretizingtheErlangdistributiongc(z):g(y)=1Xw=1PE(w)(y)]TJ /F3 11.955 Tf 11.96 0 Td[(w), (4)withPE(w)=P(w)]TJ /F6 11.955 Tf 11.95 0 Td[(1><>>:PB(0)forw=0P1j=1P(j)e(w)PB(j)forw>0. (4) 75

PAGE 76

Here,Pje(w)istheprobabilitytohavewerrorsinjburstsintheintervalofNbits:P(j)e(w)=8>><>>:Pwn=1P(j)]TJ /F5 7.97 Tf 6.59 0 Td[(1)e(w)]TJ /F3 11.955 Tf 11.95 0 Td[(n)PE(n)forj>1PE(w)forj=1, (4)andPB(j)representstheprobabilityofhavingjburstsinaninterval:PB(j)=j j!e)]TJ /F9 7.97 Tf 6.59 0 Td[(. (4)From( 4 ),( 4 ),( 4 )and( 4 ),weknowthatthecomputationofPN(w)reliesonthevaluesofand.Nowweshouldndawaytoobtainthevaluesofthesetwoparameters.WedenotethemeanvalueoferrorsinNbits(i.e.,thevalueofEfg(y)gwhenthereareNbits)asNe.IfweknowNe,wecancomputethevalueoffrom( 4 ).Accordingto[ 13 ],thevalueofdependsontheprobabilitythataburstoccursandcauseserrorsintheintervalofNbits.Itisgivenby:=Nr Ne, (4)whererisaparametercalledBitErrorRate.Finally,thevalueofNeisevaluatedasfollows:Ne=NNm p0(N+Lm)]TJ /F6 11.955 Tf 11.96 0 Td[(1), (4)wherep0istheprobabilityofhavingatleastoneerrorintheconsideredbitswhenaburstoccurs,andp0isgivenby:p0=1)]TJ /F6 11.955 Tf 11.95 0 Td[((1)]TJ /F3 11.955 Tf 34.75 8.09 Td[(Nm N+Lm)]TJ /F6 11.955 Tf 11.95 0 Td[(1)N. (4)Here,NmisthemeanvaluefornumberoferrorsperburstandLmisthemeanvalueofbursterrorlength. 76

PAGE 77

From( 4 ),( 4 )and( 4 ),wecanseethatthecomputationofanddependsonthevaluesofNm,Lm,randN,whichareinputsystemparameters.Aftercomputingthevaluesofand,wecanobtainthevalueofPN(w).Forexample,ifNm=9.5,Lm=33.5,r=10)]TJ /F5 7.97 Tf 6.59 0 Td[(3andN=10,thevalueofandarerespectively0.52and4.110)]TJ /F5 7.97 Tf 6.58 0 Td[(3.Then,PN(w)canbecalculatedby( 4 ).FromSection 4.4 ,weknowthateachslotonlycarriesbinaryinformationofeither'1'or'0'.Theinformationinthetimeframethereforerepresentsabitarrayoflengthf.Asthetimeframeisdividedintosub-frames,eachcontaininglslots,wealsodividethebitarrayintosegmentsoflbits,whichallowsourprotocoltodealwithonesegmentatatime.Considerthejthsub-frameFj.Recallfrom( 4 )thattheprobabilityofhavingw(0wl)errorsinlbitsisPl(w).Then,theprobabilityforamissingtag,whichisassignedtoFj,tobedetectedispPk+1Pl(w)(1)]TJ /F7 7.97 Tf 13.25 4.71 Td[(w l).ThedetectionprobabilityofMSMDunderthebursterrormodelcanbecomputedbysummingoverallpossiblevaluesofw.Hence,thedetectionprobabilityofMSMDunderthebursterrormodel,denotedasPmsmd)]TJ /F7 7.97 Tf 6.59 0 Td[(be,isPmsmd)]TJ /F7 7.97 Tf 6.58 0 Td[(be(p,f,Nm,Lm,r)=1)]TJ /F6 11.955 Tf 11.95 0 Td[((1)]TJ /F7 7.97 Tf 19.51 14.94 Td[(lXw=0pPk+1Pl(w)(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(w l))m. (4)Underthismodel,wecanobtainthevaluesofpopt,f(popt),pt,andfoptbyfollowingthesameprocedureasdescribedinSection 4.5.6 ,exceptthat( 4 )isreplacedwith( 4 ). 4.7NumericalResultsWehaveperformedextensivesimulationstostudytheperformanceoftheproposedMSMD,MSMD-reandMSMD-be,andcompareitwithEMDandTRP[ 46 ].Thedesignofallvesprotocolsensurethatthedetectionrequirementspeciedbymandisalwaysmet.Thisisindeedwhatweobserveinoursimulations. 77

PAGE 78

Theperformancecomparisonismadeintermsofenergyefciencyandtimeefciency,givenacertaindetectionrequirement.Tomeasuretheprotocolexecutiontime,wesetthetransmissionparametersbasedonthetypicalsettingoftheEPCglobalGen-2standard[ 14 ].Anytwoconsecutivetransmissions(fromthereadertotagsorviceversa)areseparatedbyawaitingtimeof266.4s.Thetransmissionratefromthereadertotagsis40.97Kb/sec;ittakes24.41sforthereadertotransmitonebit.A96-bitslotthatcarriesaseed-selectionsegmentis2609.76slong,whichincludesawaitingtimebeforethetransmission.Thetransmissionratefromatagtothereaderisalso40.97Kb/sec,sothatasingle-bitslotTshortforatagtorespond(i.e.,makethechannelbusy)is290.81s,alsoincludingawaitingtime.Foreachsetofsystemparameters,includingm,,andn,TRPwillcomputeitsoptimalframesize.Oncetheframesizefisdetermined,theexecutiontimeisknown,whichisfTshortplusthetimeforbroadcastingadetectionrequest.MSMD,MSMD-re,MSMD-beandEMDwillchooseasamplingprobabilityp,andcomputetheoptimalframesizefunderthatsamplingprobability.SinceEMDisaspecialcaseofMSMD,weuseouranalyticalframeworkintheprevioussectiontocomputeit.ForEMD,itsexecutiontimeisfTshort,plusthetimeforarequest.ForMSMD,MSMD-reandMSMD-be,weneedtoaddthetimefortransmittingtheselectionvector.Wecannotndawell-acceptedenergymodelforRFIDtagsordetailedparametersofenergyexpenditureforaRFIDstandard.However,aswehaveexplaininSection 4.5.1 ,theenergycostcanbeindirectlymeasuredbythenumberofparticipatingtagsbecausetheformerisproportionaltothelatter.Weusethismeasurementtostudytheenergy-timetradeoffinrelativetermsandmakeperformancecomparison.Aspartofourfuturework,wewillinvestigatetheexactenergycost(inmJ)oftagswhenanappropriateenergymodelforRFIDtagsbecomesavailable.Althoughtheexactenergyconsumptionofatagbesimulatedatthistime(whichdependsonphysical-layer 78

PAGE 79

implementation),wepointoutthateachparticipatingtagonlyneedstoreceiveasmallamountofdatafromthereaderandtransmitsone-bitinformation. 4.7.1Energy-TimeTradeoffLetn=50,000,=95%,andm=50.ForMSMD-be,therequiredinputparametersettingissimilartothatin[25]:Nm=9.5,Lm=33.5andr=10)]TJ /F5 7.97 Tf 6.59 0 Td[(3.Fig. 4-15 showstheenergy-timetradeoffcurvesproducedbysimulations.RecallthattheenergycostofMSMDorEMDisproportionaltop.Thepointatp=1ontheEMDcurverepresentsTPR.Clearly,MSMDsignicantlyoutperformsEMD.MSMDwithk=7usesthree-bitelementsintheselectionvector,whileMSMDwithk=3usestwo-bitelements.Eventhoughitincursmoreoverheadintheselectionvector,MSMDwithk=7slightlyoutperformsMSMDwithk=3.Furtherincreasingkcannotbringperformancegainduetooverlylargeoverheadfortheselectionvector.Furthermore,itisalsoshowninFig. 4-15 thatwhenwetaketheimpactofchannelerrorintoconsideration,theperformanceofMSMDwithk=7degradesslightly,whichcanbeillustratedbythecurvesofMSMD-reandMSMD-be.ForMSMD-re,increasingerrwillcausemoreexecutiontimeandenergycostbecauseoflargeprobabilityforaslottoincurerror.Eventhoughchannelerrorscausemoreoverheadinmissing-tagdetection(whichshouldbeanexpectedconsequence),akeyndingisthatthegeneralenergy-timetradeoffrelationstaythesame.InFig. 4-16 ,wezoominforadetailedlookatthecurvesegmentinthesamplingprobabilityrangeof[0,0.2].Whenp=0.08,theexecutiontimeofMSMDwithk=7is11.2%ofthetimetakenbyEMD.Whenwextheexecutiontimeat5seconds,thenumberofparticipatingtagsinMSMDwithk=7is46.7%ofthenumberinEMD.WedonotdirectlycompareMSMD-reandMSMD-dewithEMDbecausethelatterdoesnotconsiderchannelerror.Wevarythevaluesofn,andm.Similarconclusionscanbedrawnfromthesimulationresults.ThetradeoffcurvesinFig. 4-16 agreewithouranalyticalresultsinFig. 4-14 inprinciple.Wewanttostressthatoursimulationsdonotsimplyreproducetheanalytical 79

PAGE 80

Figure4-15. Protocolexecutiontimewithrespecttosamplingprobability,when=95%,m=50,andn=50,000. Figure4-16. Zoom-inviewofenergy-timetradeoffinFigure 4-15 inthesamplingprobabilityrangeof[0,0.2]. results.SimulationsconsidersystemdetailsbyusingarealRFIDspecication.Suchdetailsarenotcapturedbyanalysis.Inaddition,simulationsconsidertheexactimpactofselectionvectoronexecutiontime(measuredinseconds),insteadofcharacterizingtimeinanindirectwayusingtheframesize.InTable 5-3 ,weshowtherelativeperformanceofMSMD(k=7)withrespecttoTRP,wheren=50,000,=95%,andm=50,100,or200.MSMDisoperatedundersamplingprobabilityptandpopt.Forexample,whenm=50,pt=0.085andpopt=0.055.ThenumbersinthetableareratiosofMSMD'senergycost(orexecutiontime)toTRP'senergycost(orexecutiontime).Forexample,whenm=200,theenergycostofMSMDwithptis2.1%ofwhatTRPconsumes,anditsexecutiontimeis4.09%ofthetimeTRPtakes.AgainwedonotdirectlycompareMSMD-reandMSMD-dewithTRPbecausethelatterdoesnotconsiderchannelerror. 80

PAGE 81

Table4-3. RelativeenergycostandexecutiontimeofMSMD(k=7)underpoptandpt,when=95%andn=50,000 ptpoptenergytimeenergytime m=2002.1%4.09%1.4%269.1%m=1003.6%5.61%2.5%226.2%m=508.1%10.84%5.5%82.3% 4.7.2PerformanceComparisonNext,wecomparetheperformanceofMSMD(k=7),MSMD-re(err=5%andk=7),MSMD-be(k=7),EMD,andTRPunderdifferentvaluesofm,andn.MSMD,MSMD-re,MSMD-beandEMDareoperatedwiththeiroptimalsamplingprobabilitiespt.InFig. 4-17 4-19 ,wekeepm=50andvarythevalueof.InFig. 4-17 ,welet=99.9%,meaningthateachprotocolexecutionshoulddetectanymissing-tageventwithprobability99.9%.Theleftplotcomparestheenergycostofveprotocolswithrespectton,andtherightplotcomparestheirexecutiontimes.MSMDhasasmallerenergycostthanEMD,whichinturnhasamuchsmallerenergycostthanTRP.Takingtheimpactofchannelerrorintoconsideration,theenergycostsbyMSMD-beandMSMD-reincreasemodestlyoverMSMD.Inthemeanwhile,MSMDalsohasamuchsmallerexecutiontimethanEMDandTRP.Takingchannelerrorintoconsideration,theexecutiontimesofMSMD-beandMSMD-reincreasemodestlyoverMSMD.SimilarresultscanbedrawnfromFig. 4-18 where=99%andFig. 4-19 where=90%.Inthelattercase,theexecutiontimeofMSMDislessthanasecond.InFig. 4-20 4-21 ,wekeep=99%andvarythevalueofm.InFig. 4-20 ,m=25.InFig. 4-21 ,m=100.TheperformanceofMSMDremainsthebestamongallve. 4.8SummaryThisdissertationproposesanewprotocoldesignthatintegratesenergyefciencyandtimeefciencyformissing-tagdetection.ItusesmultiplehashseedstoprovidemultipledegreesoffreedomfortheRFIDreadertoassigntagstosingletonslots,during 81

PAGE 82

Figure4-17. A)Thenumberofparticipatingtagswithrespecttothenumberoftags,whenm=50and=99.9%.B)Theprotocolexecutiontimewithrespecttothenumberoftags,whenm=50and=99.9%. Figure4-18. SameasthecaptionofFig. 4-17 exceptfor=99%. Figure4-19. SameasthecaptionofFig. 4-17 exceptfor=90%. Figure4-20. A)Thenumberofparticipatingtagswithrespecttothenumberoftags,whenm=25and=99%.B)Theprotocolexecutiontimewithrespecttothenumberoftags,whenm=25and=99%. Figure4-21. SameasthecaptionofFig. 4-20 exceptform=100. 82

PAGE 83

whichthetagsannouncetheirpresenceintheprocessofmissing-tagdetection.Werstpresentthisnewprotocolwithreliablechannels.Theresultisamulti-foldcutinbothenergycostandexecutiontime.Suchperformanceimprovementiscriticalforaprotocolthatneedstobeexecutedfrequently.Then,weextendtheprotocoltoconsidertwocategoriesofchannelerrorsinducedbynoise/interferenceintheenvironment.Theinvolvingofchannelerrorswillmaketheenergy/timegainsslightlyreduced,butremainsignicantcomparingtoEMDandTRP.Wealsorevealafundamentalenergy-timetradeoffintheprotocoldesign.Thistradeoffgivesexibilityinperformancetuningwhentheprotocolisappliedinpracticalenvironment. 83

PAGE 84

CHAPTER5ANEFFICIENTPROTOCOLFORRFIDMULTIGROUPTHRESHOLD-BASEDCLASSIFICATIONThischapterstudiestheRFIDmultigroupthreshold-basedclassicationproblem.Thegoalistodeterminewhetherthenumberofobjectsineachgroupisaboveorbelowaprescribedthresholdvalue.Solvingthisproblemisimportantforinventorytrackingapplications.Ifthenumberofgroupsisverylarge,itwillbeinefcienttomeasurethegroupsoneatatime.Thebestexistingsolutionformultigroupthreshold-basedclassicationisbasedongenericgrouptesting,whosedesignishowevergearedtowardsdetectingasmallnumberofpopulousgroups.Itsperformancedegradesquicklywhenthenumberofgroupsabovethethresholdbecomelarge.Inthisdissertation,weproposeanewclassicationprotocolbasedontagsamplingandlogicalbitmaps.Itachieveshighefciencybymeasuringallgroupsinamixedfashion.Inthemeantime,weshowthatthenewmethodisabletoperformthreshold-basedclassicationwithanaccuracythatcanbepre-settoanydesirablelevel,allowingtradeoffbetweentimeefciencyandaccuracy.Therestofthechapterisorganizedasfollows.Section 5.1 presentsthesystemmodelanddenestheproblemtobesolved.Section 5.2 discussestherelatedworkandgivesthemotivationforoursolution.Section 5.3 proposesourtwo-phaseprotocolfortheRFIDthreshold-basedclassicationproblem.Section 5.4 evaluatesthenewprotocolthroughsimulations.Section 5.5 drawstheconclusion. 5.1ProblemDenitionandSystemModel 5.1.1SystemModelTherearethreetypesofRFIDtags.Passivetagsaremostwidelydeployedtoday.Theyarecheap,butdonothaveinternalpowersources.PassivetagsrelyonradiowavesemittedfromanRFIDreadertopowertheircircuitandtransmitinformationbacktothereaderthroughbackscattering.Theyhaveshortoperationalranges,typicallyafewmetersinanindoorenvironment.Tocoveralargearea,arraysofRFIDreaderantennas 84

PAGE 85

mustbeinstalled.Semi-passivetagscarrybatteriestopowertheircircuit,butstillrelyonbackscatteringtotransmitinformation.Activetagsusetheirownbatterypowertotransmit,andconsequentlydonotneedanyenergysupplyfromthereader.Activetagsoperateatamuchlongerdistance,makingthemparticularlysuitableforapplicationsthatcoveralargearea,whereoneorafewRFIDreadersareinstalledtoaccessalltaggedobjectsandperformmanagementfunctionsautomatically.Withricheronboardresources,activetagsarelikelytogainmorepopularityinthefuture,particularlywhentheirpricesdropovertimeasmanufacturaltechnologiesareimprovedandmarketsareexpanded.Communicationbetweenreadersandtagsistime-slotted.Readerssendoutarequest,whichisfollowedbyaslottedtimeframeduringwhichtagstransmitintheirselectedslots.Thereadersmaytaketurnstotransmittherequestinordertoavoidinterference,oramoresophisticatedschedulingalgorithmmaybeusedtoallowreadersthatdonotinterferetotransmitsimultaneously.Whenatagtransmits,aslongasonereaderreceivesthetransmissioncorrectly,thetransmissionwillbesuccessful.Inourprotocoldesign,wecanlogicallytreatallreadersasone,whichtransmitsarequestandthenlistenstothetags'responses.Weusetwotypesofslotstocarrytagresponses.Thersttypeiscalledalong-responseslot,whoselengthisdenotedasTlong,duringwhichatagtransmitsmultiplebits,allowingthereadertotellwhetherthereiscollisioninaslot.Thesecondtypeiscalledashort-responseslot,whoselengthisdenotedasTshort,whichcarriesone-bitinformation:`0'foranemptyslotwhennotagtransmits,and`1'forannon-emptyslotwhenoneormoretagstransmitsignaltomakethechannelbusy. 5.1.2MultigroupThreshold-BasedClassicationProblemConsiderabigwarehousewithtensofthousandsofitems.EachitemisattachedwithanRFIDtagforcommunicationwithanRFIDreader.Theitemsaredividedintodifferentgroupsbasedoncertainproperties,whichcanbetheproductsub-category, 85

PAGE 86

productiondate,orproductionplace.Tosupportgrouping,eachtagIDshouldcontaintwocomponents:agroupID,whichidentiesthegroupthetagbelongsto,andamemberID,whichidentiesaspecictaginthegroup.Clearly,alltagsinagroupmustcarrythesamegroupID,whiletagsindifferentgroupscarrydifferentgroupIDs.WeassumethattheRFIDreaderknowsthegroupIDsinthesystem.Wedenethepopulationorsizeofagroupasthenumberoftagsinthisgroup.Aswehaveexplainedintheintroduction,whileitispossibletoperformprecisemultigroupclassicationathighcost,thefocusofthisdissertationistostudyefcientsolutionsforapproximatemultigroupclassication.Weformallydenetheproblemasfollows:Lethbethethresholdand1bealargeprobabilityvalue.Werequirethatanygroupwhosepopulationexceedshshouldbereportedwithaprobabilityofatleast1.Letlbeanotherintegerparametersmallerthanhand1beasmallprobabilityvalue.Wealsorequirethattheprobabilityofreportinganygroupwithlorfewertagsshouldbenomorethan1.Letk1bethepopulationofanarbitrarygroupg.Ourperformanceobjectivescanbeexpressedintermsofconditionalprobabilitiesasfollows: Probfgroupgisreportedbythereaderjk1hg1Probfgroupgisreportedbythereaderjk1lg1(5)Wetreatthereportofagroupwithlorfewertagsasafalsepositive,andthenon-reportofagroupwithhormoretagsasafalsenegative.Hence,theaboveobjectivescanalsobestatedasboundingthefalsepositiveratioby1andthefalsenegativeratioby1)]TJ /F4 11.955 Tf 11.96 0 Td[(1. 5.2Preliminary 5.2.1PriorWorkSheng,TanandListudiedthemultigroupthreshold-basedclassicationproblemin[ 45 ].Theybeginwithasimplethresholdcheckingscheme(TCS)toapproximatelyanswerswhetherthenumberoftagsexceedsathreshold.BasedonTCS,theypropose 86

PAGE 87

Figure5-1. TheestimationtimewithrespecttothegroupsizeforUPE,EZB,EnhancedFNEB,EMLEAandART,when01=99%and01=1%. Table5-1. Notations SymbolsDescriptions 1)]TJ /F4 11.955 Tf 11.95 0 Td[(1upperboundoffalsenegativeratio1upperboundoffalsepositiveratiom1bitlengthoflogicalbitmapn1numberoftagsSnumberofabove-thresholdgroupsk1actualnumberoftagsinanarbitrarygroup^k1estimatednumberoftagsinanarbitrarygrouprirandomnumberintheithpollingf1lengthofthetimeframeeachpollingH()hashfunctionwhoserangeis[0,f1)]TJ /F6 11.955 Tf 11.95 0 Td[(1]midatag'smemberIDgidatag'sgroupIDhaprescribedhigherboundthresholdvaluelaprescribedlowerboundthresholdvaluewnumberofpollings twoprobabilisticprotocols.Therstoneisbasedongenericgrouptesting(GT),whichconsistsofmultiplerounds.Ineachround,thereadershufesallgroupsintodifferentcategories,eachofwhichmaycontaintagsfrommultiplegroups.TCSisthenappliedtocheckthenumberoftagsineachcategory.Thecategorieswithsufcienttagsarelabeledaspotentialpopulouscategories,whichmayincludeabove-thresholdgroups.Intheend,thetestinghistoryisusedtoclassifyallabove-thresholdgroups.Thesecond 87

PAGE 88

protocolisacombinationofgrouptestinganddivide-and-conquer,whichignoresthecategoriesthatfailtopasstheTCStestsinthepreviousround,dividestheremainingcategoriesintomultiplesub-categories,andappliesTCStoeachsub-categoriesintheremainingrounds.Anotherpossiblesolutionforthemultigroupthreshold-basedclassicationproblemistouseareadertocollecttheactualtagIDsfromtags[ 8 12 17 20 32 34 49 54 ],whereeachIDcontainsbitsthatidentifythegroupofthetag.Applyingtotheprobleminthisdissertation,theseID-collectionprotocolsdonotworkwellforlarge-scaleRFIDsystemsduetotheirlongidenticationtime.ManymethodswereproposedtoestimatethewholepopulationofanRFIDsystem.Theyareessentiallysingle-groupestimators.Wecanusethemtorstestimateindividualgroupsizes(onegroupatatime)andthenusethesizesforclassicationpurpose.KodialamandNandagopalproposetherstsetofsingle-groupestimators,includingtheZeroEstimator(ZE),theCollisionEstimator(CE),andtheUniedProbabilisticEstimator(UPE),whichcollectinformationfromtagsinaseriesoftimeframesandestimatesthewholepopulationoftagsinthesystembasedonthenumberofemptyslotsand/orthenumberofcollisionslots[ 21 ].Afollow-upworkbythesameauthorsproposestheEnhancedZero-BasedEstimator(EZB)[ 22 ],whichisanasymptoticallyunbiasedestimatorandmakesestimationonlybasedonthenumberofemptyslots.Qianetal.provideareplicate-insensitiveestimationalgorithmcalledtheLottery-Framescheme(LoF)[ 37 ].TheEnhancedFirstNon-EmptyslotsBasedEstimator(EnhancedFNEB)[ 16 ]canbeusedtoestimatetagpopulationinbothstaticanddynamicenvironmentsbymeasuringthepositionoftherstnon-emptyslotineachframe.Lietal.[ 26 ]studytheestimationproblemforlarge-scaleRFIDsystemsfromtheenergyanglebasedonaEnhancedMaximumLikelihoodEstimationAlgorithm(EMLEA).Theydesignseveralenergy-efcientprobabilisticalgorithmsthatiterativelyreneacontrolparametertooptimizetheinformationcarriedinthe 88

PAGE 89

transmissionsfromtags,suchthatboththenumberandthesizeofthetransmissionsareminimized.TheAverageRunbasedTagestimation(ART)scheme[ 43 ]furtherreducestheexecutiontimeforpopulationestimation,basedontheaveragerunlengthofonesinthebitstringreceivedinthestandardizedframe-slottedAlohaprotocol.Finally,theZero-OneEstimator(ZOE)[ 59 ]providesfastandreliablecardinalitybytuningthesystemparametersandconvergingtotheoptimalsettingsthroughabisectionsearch. 5.2.2MotivationWerstshowtheperformanceofsomeexistingsingle-groupestimatorsthroughsimulation.Fromthesimulationresults,wearguethattheseestimatorsarenottime-efcientwhentheyareappliedtothemultigroupthreshold-basedclassicationproblem.Figure 5-1 presentstheexecutiontimeofveexistingsingle-groupestimators[ 16 21 22 26 43 ]withrespecttothenumberoftagsinthegroup;detailsaboutthesimulationsettingandparameterscanbefoundinSection 5.4.1 .Whiletheseestimatorsaredesignedtomeasurethesizeofasinglegroup,theymaybeappliedtoperformingmultigroupthreshold-basedclassicationbyestimatingonegroupatatime.Theirestimationaccuracyisspeciedbyacondenceinterval:Theprobabilityfortheestimatetodeviatefromthetruegroupsizeby01percentageormoreshouldnotexceed01,where01and01aretwopre-speciedsystemparameters.Theyaresetto99%and1%respectivelyinoursimulation.Fromthegure,weobservethattheestimationtimechangesverylittlewithrespecttothenumberoftags.Forexample,ARTtakesabout10secondstoestimatethetagpopulationintherangefrom500to50,000.Iftherearetwogroupsof25,000tagseach,thetotalestimationtimeforthetwogroupswillbe20seconds.However,ifthereare100groupsof500tagseach,thetotalestimationtimewillbe1,000seconds!Hence,theseestimatorsarenotsuitablewhentherearenumeroussmallgroups. 89

PAGE 90

ThegrouptestingmethodinGT[ 45 ]cansignicantlyreducetheexecutiontimeforpopulousgroupdiscovery.However,simulationresultsshowthattheirexecutiontimeisapproximatelyproportionaltothenumberofgroupsabovethethreshold.Hence,theperformanceoftheprotocolwilldeteriorateifthenumberofgroupsabovethethresholdislarge.Inaddition,theRFIDreadermustbeabletodistinguishthreetypesofslots:emptyslot,duringwhichnotagtransmits;singletonslot,duringwhichonlyonetagtransmits,andcollisionslot,duringwhichmorethanonetagtransmits.Wefollowtwogeneraldesignprincipleswhendesigningourtime-efcientclassicationprotocol.First,wewanttominimizethelengthofeachtimeslot.BasedontheparametersoftheEPCglobalGen-2standard[ 14 ],inordertotransmita96-bitIDfromatagtoanRFIDreader,weneedaslotof2608.8s.However,ifthereaderisnotinterestedinIDsbutwantstodistinguishcollisionslotsfromsingletonoremptyslots[ 45 ],tagsshouldtransmit10-bitlongresponses,usingslotsof470.5seach.Furthermore,ifthereaderdoesnotneedtodistinguishcollisionslotsfromsingletonslotsbutonlywantstoknowwhethertheslotsareemptyornot,tagscantransmitone-bitshortresponses,usingslotsof290.8seachtocarryonebitinformation(channelbusyoridle);thisisthetypeofslotswewilluseinourprotocoldesign.Notethat266.4swaitingtimeisincludedineachslottoseparateitfromneighboringslots.Second,wewanttominimizethenumberofslots.Figure 5-1 clearlyshowsthattraditionalapproachesofmeasuringonegroupatatimewillnotworkwellwhentherearealargenumberofgroups.Wetakeanewdesignthatisdrasticallydifferentfromtraditionalones:measuringthegroupsallatonceandprobabilisticallysharingeachslotbymultiplegroups.Thisdesignhasaninterestingfeaturethatitsexecutiontimeislargelyinsensitivetothenumberofgroupsifthetotalnumberoftagsisaboutthesame.Itmakesourprotocolparticularlysuitableforsituationswheretherearealargenumberofsmallormedium-sizedgroups. 90

PAGE 91

5.3AnEfcientThreshold-BasedClassicationProtocolThissectionpresentsanefcientThreshold-BasedClassication(TBC)Protocol,whichisacombinationofdynamicslotsharingamonggroupsandmaximumlikelihoodestimationofgroupsizes. 5.3.1DynamicSlotSharingWeshareallslotsamongallgroups.Notably,weabandontheapproachofapplyingasingle-groupestimator[ 16 21 22 26 45 ]tomeasureonegroupatatime,butinsteadmeasurethesizesofallgroupstogetherinonetimeframewhoseslotsareshared:EachgroupIDispseudo-randomlyhashedtoacertainnumberofslotsinthetimeframe.Eachtaginthegroupwillprobabilisticallypickoneoftheseslotstotransmit.Listeningtothechannel,thereaderconvertsthetimeframeintoabitmap.Foreachgroup,itextractsthebitsthatthegroupIDishashedto.Thosebitsformthelogicalbitmapofthegroup,fromwhichthegroupsizeisestimated.Inthisapproach,eachbitandthecorrespondingslotmaybesharedbymorethanonegroup.Thissharingintroducesnoise;thelogicalbitmapofonegroupmaycarrysomebitsthataresetto`1'notbytransmissionsoftagsinthisgroup,butbytransmissionsoftagsfromothergroupsthathappentobehashedtothesametimeslots.Fortunately,inabird's-eyeview,allslotsaresharedbyallgroupsuniformlyatrandom(throughindependenthashing),whichmeansthenoiseisuniformlydistributedinthewholetimeframe.Suchuniformnoiseismeasurable.Toestimatethesizeofagroup,wewillusethelogicalbitmapofthatgroup,butsubtractthenoisethattagsfromothergroupsintroduce.Tofurtherimproveperformance,thereaderrepeatstheaboveapproachmultipletimestogathermultipleindependentlogicalbitmapsforeachgroup,andestimationbasedonmultiplelogicalbitmapsreducesthevarianceoftheresult.Sharingslotssavestime.Forexample,ifweshareeachslotamong30groupsonaverage(asweobserveinatypicalsimulationofSection 5.4 ),wewillbeabletoachieveafactorof30reductioninexecutiontime.However,sharingslotscausenoiseamong 91

PAGE 92

groupsduringsizeestimation.Althoughthenoiseisstatisticallyuniformlydistributed,itsvariancerequiresustorepeatforadditionallogicalbitmapsinordertoaverageoutthenoisevariance,whichmeanslongexecutiontime.Fortunately,thetimesavedbysharingoutweightsthetimeneededfornoiseremovalaswewilldemonstratelater. 5.3.2OverviewOurthreshold-basedclassication(TBC)protocolconsistsofthreephases:theparameter-precomputingphase,theframephase,andthereportphase.Theparameter-precomputingphasecomputessystemparametersforoptimalperformanceoftheprotocol.Usingtheseparameters,theframephasemakesw(1)pollingrequests,eachofthemfollowedbyatimeframe,duringwhichtagsofallgroupstransmitinselectedslots.Thereaderconvertseachtimeframeintoabitmap,fromwhichlogicalbitmapsareextracted.Usingtheselogicalbitmaps,thereportphaseemploystheMaximumLikelihoodEstimation(MLE)methodtoreporttheabove-thresholdgroups.Therearefoursystemparameters:wisthenumberofpollings(ortimeframes),p1isasamplingprobability,f1isthesizeofeachtimeframe,i.e.,thenumberofslotsinaframeorthenumberofbitsinthebitmapthattheframeisconvertedto,andm1isthesizeofeachlogicalbitmap.Clearly,m1
PAGE 93

5.3.3FramePhaseTheframephaseiscomposedofwpollings,whichareperformedinasimilarway:Intheithpolling(where1iw),anRFIDreaderrstbroadcastsarequestmessage,includingarandomnumberrandthesystemparameters,p1,f1andm1.Therequestmessagealsoservesthepurposeofsynchronizingtheclocksofalltagsforstartingatimeframeoff1slotsrightaftertherequest.Consideranarbitrarytagtinanarbitrarygroupg.Thetagdecideswithaprobabilityp1forwhethertoparticipateinthecurrentpolling.Ifitdecidesnotto,itwillkeepsilentuntilthenextpollingrequest.Ifthetagdecidestoparticipate,itcomputesahashvalueH(gidLF(ri,H(tid)modm1))astheindexofthetimeslotselectedforitstransmission,whereH()isahashfunctionwhoserangeis[0,f1)]TJ /F6 11.955 Tf 12.6 0 Td[(1],gidisthetag'sgroupID,tidisthetag'smemberID,andF(x,y)isapseudo-randomnumbergeneratorthattakestwoinputparameters:xandy.F(x,y)usesxastheseed,generatesyrandomnumbers,andoutputstheythnumber.ThetransmissionsfromallparticipatingtagsformabitmapBi.Clearly,fortagsofgroupg,theindicesoftheirselectedslotsintheframecanonlybeH(gidLF(ri,0)),H(gidLF(ri,1)),...,H(gidLF(ri,m1)]TJ /F6 11.955 Tf 12.28 0 Td[(1)).Theseslotsormoreprecisely,thebitsconvertedfromtheseslots,formthelogicalbitmapofg,denotedasLBi(g).NotethatthevalueofH(tid)modm1givestheindexofthecorrespondingbitinthelogicalbitmap.Forexample,ifatagselectstheH(gidLF(ri,H(tid)modm1))thslottotransmit,the(H(tid)modm1)thbitinthelogicalbitmapwillbesetto`1'.Essentially,weembedthelogicalbitmapsofallinBi.WepointoutthatthecomplexityofourprotocolismostlyplacedattheRFIDreader,whichhastocomputetheoptimalsystemparameters,initiatetheprotocol,receivetagtransmissions,andperformclassication(seethenexttwosubsections).Thetag'soperationisrelativelysimple:receivingarequestfromthereader,performinghash, 93

PAGE 94

andtransmittinginatimeslot.TocomputeH(gidLF(ri,H(tid)modm1)),weexpecttagstoimplementapseudo-randomnumbergeneratorFasrequiredby[ 14 ].AhashfunctionmaybeimplementedfromFbyusingthehashinputastheinputtoF.Thereareothersimplewaysofimplementingahashfunctionfortags,suchas[ 24 ],whichusesapre-storedbitringtoproducehashoutput.Ourprotocolcannotbedirectlysupportedbytoday'soff-the-shelfC1G2-compatibletagsbecausethecurrentstandarddoesnotsupportoperationsongroupIDs(suchashashingbasedonagroupIDfortheindexofaslot).However,webelievefuturetags(orstandards)maybeenhancedtosupportsuchaprotocol.Inourcase,anewoperationalcodecalledClassicationneedstobedened,groupIDsneedtobestandardized,andoperationsbasedongroupIDs(suchashashing)needtobeimplementedontags. 5.3.4ReportPhaseAfterwpollings,thereaderobtainswbitmaps,Bi,1iw.Itsendsthebitmapstoanofinedataprocessingmodule.There,thelogicalbitmapsofeachgroupisextracted.Foranarbitrarygroupg,weextractalogicalbitmapLBi(g)fromBiasfollows:SetthejthbitofLBi(g)tobetheH(gidLF(ri,j))thbitinBi,i.e.,LBi(g)[j]=Bi[H(gidLF(ri,j))],where1iwand0jm1)]TJ /F6 11.955 Tf 11.96 0 Td[(1.LetxibethenumberofzerosobservedinLBi(g).Letn1bethetotalnumberoftagsinthesystemandk1betheactualpopulationofgroupg.Belowwederivetheformulatocomputeanestimate^k1ofthepopulation.ConsidertheithpollingintheframephaseandanarbitrarybitbinLBi(g).Atagingroupghasaprobabilityofp1 m1toselectthisbitandsetitto`1'becausethetagissampledwithprobabilityp1andifsampled,itonlysetsoneofthem1bitsinthelogicalbitmapofg.Anytaginothergroupshasaprobabilityofp1 f1tosetthisbitto`1'duetodynamicslotsharingacrossthewholeframe.Hence,theprobabilityforbtoremainzerois q=(1)]TJ /F3 11.955 Tf 13.15 8.08 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F3 11.955 Tf 14.9 8.08 Td[(p1 m1)k1.(5) 94

PAGE 95

Hence,thelikelihoodfunctionLiforustoobservexibitsofzerosinLBi(g)isLi=((1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F3 11.955 Tf 14.9 8.09 Td[(p1 m1)k1)xi(1)]TJ /F6 11.955 Tf 11.95 0 Td[((1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.58 0 Td[(k1(1)]TJ /F3 11.955 Tf 14.89 8.09 Td[(p1 m1)k1)m1)]TJ /F7 7.97 Tf 6.58 0 Td[(xi. (5)ThelikelihoodfunctionLforustoobserveallxivalues,1iw,inthewlogicalbitmapsisL=Qwi=1Li.Thatis,L=wYj=1((1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F3 11.955 Tf 14.89 8.09 Td[(p1 m1)k1)xi(1)]TJ /F6 11.955 Tf 11.96 0 Td[((1)]TJ /F3 11.955 Tf 13.15 8.08 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.58 0 Td[(k1(1)]TJ /F3 11.955 Tf 14.9 8.08 Td[(p1 m1)k1)m1)]TJ /F7 7.97 Tf 6.59 0 Td[(xi. (5)Wewanttondanestimate^k1thatmaximizesL,namely, ^k1=argmaxfLgk1.(5)Sincethemaximumisnotaffectedbymonotonetransformations,wetakethelogarithmofbothsidesof( 5 ):ln(L)=wXi=1xi((n1)]TJ /F3 11.955 Tf 11.96 0 Td[(k1)ln(1)]TJ /F3 11.955 Tf 13.15 8.08 Td[(p1 f1)+k1ln(1)]TJ /F3 11.955 Tf 14.9 8.08 Td[(p1 m1))+(m1)]TJ /F3 11.955 Tf 11.96 0 Td[(xi)ln(1)]TJ /F6 11.955 Tf 11.95 0 Td[((1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F3 11.955 Tf 14.89 8.09 Td[(p1 m1)k1)). (5)Differentiatingbothsidesoftheaboveequation,wehave@lnL @k1=wXi=1(xi)]TJ /F3 11.955 Tf 11.96 0 Td[(m1(1)]TJ /F7 7.97 Tf 13.15 5.03 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.58 0 Td[(k1(1)]TJ /F7 7.97 Tf 14.39 5.03 Td[(p1 m1)k1 1)]TJ /F6 11.955 Tf 11.95 0 Td[((1)]TJ /F7 7.97 Tf 13.15 5.04 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F7 7.97 Tf 14.38 5.04 Td[(p1 m1)k1)(ln(1)]TJ /F3 11.955 Tf 14.89 8.09 Td[(p1 m1))]TJ /F6 11.955 Tf 11.95 0 Td[(ln(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)). (5)Aftersettingtherightsidetozeroandsimplifyingit,wehavethefollowingestimator, ^k1=ln(Pwi=1xi mw(1)]TJ /F13 5.978 Tf 7.78 4.39 Td[(p1 f1)n1) ln(1)]TJ /F13 5.978 Tf 8.71 4.39 Td[(p1 m1 1)]TJ /F13 5.978 Tf 7.78 4.4 Td[(p1 f1).(5) 95

PAGE 96

In( 5 ),m1andf1aregivenparameterswhosevaluesarepre-computedbythereader.Thevaluesofxi,1im1,areobtainedfromLBi(g).Thetotalnumbern1oftagscanbeestimatedfromthebitmaps,Bi,1iw.LetXibethenumberofzerosinBi.TheprobabilityforeachtagtobesampledandsetacertainbitinBito`1'isp1 f1.Theprobabilityforeachbittoremainzeroisapproximately(1)]TJ /F7 7.97 Tf 13.33 5.04 Td[(p1 f1)n1.ThelikelihoodfunctionforustoobserveXizerosinBi,1iw,is L=wYi=1(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1Xi(1)]TJ /F6 11.955 Tf 11.95 0 Td[((1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1)f1)]TJ /F7 7.97 Tf 6.59 0 Td[(Xi(5)Usingthemaximumlikelihoodestimation,wetakethelogarithmofbothsides,differentiateit,andthenlettherightsidebezero.WehavewXi=1Xi)]TJ /F3 11.955 Tf 11.96 0 Td[(wf1(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1=0n1=lnPwi=1Xi wf ln(1)]TJ /F7 7.97 Tf 13.15 5.03 Td[(p1 f1), (5)whereXi,1iw,areobtainedfromBi.Foreachgroupg,afterweestimateitspopulation^k1basedon( 5 ),wereportthegroup(asanabove-thresholdgroup)if^k1T,whereTisanothersystemparameterthatwillbedeterminedinthenextsubsectionbasedontheprobabilisticperformanceobjectives( 5 ). 5.3.5Parameter-PrecomputingPhaseWerstdeveloptheconstraintsthatthesystemparametersmustsatisfyinordertoachievetheprobabilisticperformanceobjectives.Basedontheconstraints,wedeterminetheoptimalvaluesforthelengthm1oflogicalbitmaps,thenumberwofpollings,theframesizef1,andtheparameterT.Agroupgwhoseestimatedpopulationis^k1willbereportedif ^k1T.(5) 96

PAGE 97

Thatis,ln(Pwi=1xi mw(1)]TJ /F13 5.978 Tf 7.78 4.39 Td[(p1 f1)n1) ln(1)]TJ /F13 5.978 Tf 8.7 4.4 Td[(p1 m1 1)]TJ /F13 5.978 Tf 7.78 4.4 Td[(p1 f1)TwXi=1xiwm(1)]TJ /F7 7.97 Tf 14.38 5.03 Td[(p1 m1 1)]TJ /F7 7.97 Tf 13.15 5.04 Td[(p1 f1)T(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1. (5)LetC=wm(1)]TJ /F13 5.978 Tf 8.71 4.39 Td[(p1 m1 1)]TJ /F13 5.978 Tf 7.79 4.39 Td[(p1 f1)T(1)]TJ /F7 7.97 Tf 13.37 5.04 Td[(p1 f1)n1.Therefore,theprobabilityforthereadertoreportagroupisProb(^k1T)=Prob(Pwi=1xiC).From( 5 ),weknowthatxifollowsthebinomialdistributionwithparametersm1and(1)]TJ /F7 7.97 Tf 13.16 5.03 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F7 7.97 Tf 14.38 5.03 Td[(p1 m1)k1: xiBino(m1,(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.58 0 Td[(k1(1)]TJ /F3 11.955 Tf 14.89 8.09 Td[(p1 m1)k1).(5)SinceabinomialdistributionBino(a,b)canbeexcellentlyapproximatedbyanormaldistributionNorm(ab,ab(1)]TJ /F3 11.955 Tf 10.31 0 Td[(b))whenaislargeenough(whichisthecaseform1),( 5 )canbeapproximatelywrittenasxiNorm(m1(1)]TJ /F15 10.909 Tf 12.1 7.39 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F15 10.909 Tf 13.7 7.39 Td[(p1 m1)k1,m1(1)]TJ /F15 10.909 Tf 12.11 7.38 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F15 10.909 Tf 13.7 7.38 Td[(p1 m1)k1(1)]TJ /F17 10.909 Tf 10.9 0 Td[((1)]TJ /F15 10.909 Tf 12.11 7.38 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F15 10.909 Tf 13.69 7.38 Td[(p1 m1)k1)). (5)Accordingto( 5 ),weknowwXi=1xiNorm(mw(1)]TJ /F15 10.909 Tf 12.11 7.38 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F15 10.909 Tf 13.7 7.38 Td[(p1 m1)k1,mw(1)]TJ /F15 10.909 Tf 12.1 7.38 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.58 0 Td[(k1(1)]TJ /F15 10.909 Tf 13.69 7.38 Td[(p1 m1)k1(1)]TJ /F17 10.909 Tf 10.91 0 Td[((1)]TJ /F15 10.909 Tf 12.1 7.38 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.58 0 Td[(k1(1)]TJ /F15 10.909 Tf 13.69 7.38 Td[(p1 m1)k1)). (5)Let=mw(1)]TJ /F7 7.97 Tf 12.24 5.03 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F7 7.97 Tf 13.48 5.03 Td[(p1 m1)k1and2=mw(1)]TJ /F7 7.97 Tf 12.24 5.03 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(k1(1)]TJ /F7 7.97 Tf 13.48 5.03 Td[(p1 m1)k1(1)]TJ /F6 11.955 Tf 11.05 0 Td[((1)]TJ /F7 7.97 Tf 12.25 5.03 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.58 0 Td[(k1(1)]TJ /F7 7.97 Tf -456.27 -18.88 Td[(p1 m1)k1),thenwehave Prob(wXi=1xi=j)=1 p 22e)]TJ /F21 5.978 Tf 5.76 0 Td[((j)]TJ /F24 5.978 Tf 5.76 0 Td[()2 22(5) 97

PAGE 98

Thus,Prob(^k1T)=Pro(wXi=1xiC)=CXj=01 p 22e)]TJ /F21 5.978 Tf 5.76 0 Td[((j)]TJ /F24 5.978 Tf 5.75 0 Td[()2 22 (5)Therstperformanceobjectivein( 5 )canbetranslatedintoProb(^k1Tjk1h)1,whichis CXj=01 p 22e)]TJ /F21 5.978 Tf 5.75 0 Td[((j)]TJ /F24 5.978 Tf 5.76 0 Td[()2 221(5)wherek1h.Sincetheleftsideoftheinequalityisanincreasingfunctionofk1,wecanreplacethetermk1withh.Then,wehavetherstconstraintforthesystemparameters. CXj=01 p 212e)]TJ /F21 5.978 Tf 5.76 0 Td[((j)]TJ /F24 5.978 Tf 5.76 0 Td[(1)2 2121,(5)where1=mw(1)]TJ /F7 7.97 Tf 13.15 5.03 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(h(1)]TJ /F7 7.97 Tf 14.38 5.03 Td[(p1 m1)hand12=1(1)]TJ /F6 11.955 Tf 11.96 0 Td[((1)]TJ /F7 7.97 Tf 13.15 5.03 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(h(1)]TJ /F7 7.97 Tf 14.38 5.03 Td[(p1 m1)h).Similarly,thesecondperformanceobjectivein( 5 )canbetranslatedintothefollowingconstraint, CXj=01 p 222e)]TJ /F21 5.978 Tf 5.76 0 Td[((j)]TJ /F24 5.978 Tf 5.75 0 Td[(2)2 2221,(5)where2=mw(1)]TJ /F7 7.97 Tf 13.15 5.04 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(l(1)]TJ /F7 7.97 Tf 14.38 5.04 Td[(p1 m1)land22=2(1)]TJ /F6 11.955 Tf 11.95 0 Td[((1)]TJ /F7 7.97 Tf 13.15 5.04 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.58 0 Td[(l(1)]TJ /F7 7.97 Tf 14.39 5.04 Td[(p1 m1)l).WewanttondoptimalsystemparametersthatminimizetheexecutiontimerequiredbyTBC,i.e.,wf1,subjecttotheabovetwoconstraints. 98

PAGE 99

Minimizewf1, (5)subjecttoCXj=01 p 212e)]TJ /F21 5.978 Tf 5.76 0 Td[((j)]TJ /F24 5.978 Tf 5.76 0 Td[(1)2 2121CXj=01 p 222e)]TJ /F21 5.978 Tf 5.76 0 Td[((j)]TJ /F24 5.978 Tf 5.76 0 Td[(2)2 2221C=mw(1)]TJ /F7 7.97 Tf 14.38 5.04 Td[(p1 m1 1)]TJ /F7 7.97 Tf 13.15 5.04 Td[(p1 f1)T(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n11=mw(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(h(1)]TJ /F3 11.955 Tf 14.89 8.09 Td[(p1 m1)h12=1(1)]TJ /F6 11.955 Tf 11.96 0 Td[((1)]TJ /F3 11.955 Tf 13.15 8.08 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(h(1)]TJ /F3 11.955 Tf 14.89 8.08 Td[(p1 m1)h)2=mw(1)]TJ /F3 11.955 Tf 13.15 8.09 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(l(1)]TJ /F3 11.955 Tf 14.89 8.09 Td[(p1 m1)l22=2(1)]TJ /F6 11.955 Tf 11.96 0 Td[((1)]TJ /F3 11.955 Tf 13.15 8.08 Td[(p1 f1)n1)]TJ /F7 7.97 Tf 6.59 0 Td[(l(1)]TJ /F3 11.955 Tf 14.89 8.08 Td[(p1 m1)l).Theparametersh,l,1and1aregivenbytheperformanceobjectives( 5 ).Tosolvetheaboveconstrainedoptimizationproblem,weneedtodeterminetheoptimalvaluesoftheremainingvesystemparametersp1,w,f1,m1andT,suchthatwf1isminimized.Wecanapproximatelysolve( 5 )throughsearchingapresetparameterspace.SeeSection 5.4 forthesearchalgorithmusedinoursimulationsandtheresultingprotocolperformance.Thevalueofn1canbeestimatedthroughanestimationprotocol[ 21 43 59 ].Moreover,oncetheperformanceobjectivesaredecided,wecanprecomputethesystemparameters(p1,w,f1,m1andT)fordifferentvaluesofn1(e.g.,atstepsof500).Afterthecurrentnumberoftagsinthewholesystemisestimated,thesystemparameterscanbequicklylookedupfromtheprecomputedresults. 5.4NumericalResults 5.4.1SettingWeevaluatetheperformanceofTBCandcompareitwiththeGroupTesting(GT)[ 45 ],theAverageRunbasedTagestimation(ART)scheme[ 43 ]andtheZero-One 99

PAGE 100

Estimator(ZOE)[ 59 ].GTisthemostrelatedwork.Itprobabilisticallyidentiespopulousgroupswhosesizesarelargerthanathreshold.WedenotetheproposedprotocolTBCwiththeoptimalsamplingprobabilityp1asTBC(p1),andTBCwithp1=100%asTBC(100%),whichisaspecialcaseofTBCwheresamplingisnotapplied,asispublishedintheconferenceversionofthisdissertation[ 30 ].Wewanttoseehowmuchimprovementcanbeachievedthroughoptimalsampling.ZOEandARTaredesignedforRFIDpopulationestimation,notforsatisfyingtheprobabilityperformanceobjectivesin(1).However,theestimationresultsfromthesetwoestimatorscanbeusedforclassicationbyreportingthosegroupswhoseestimatedsizesareaboveathreshold.Morespecically,theyestimatethegroupsizesonegroupatatime.Foreachgroup,theyprogressivelyimprovethesizeestimation^k1.Wewillrejectagroupwhen^k1h+l 2andtheprobabilityofjk1)]TJ /F6 11.955 Tf 14.27 2.66 Td[(^k1j>h)]TJ /F6 11.955 Tf 14.26 2.66 Td[(^k1isnogreaterthan1)]TJ /F4 11.955 Tf 12.02 0 Td[(1.Wewillacceptagroupwhen^k1h+l 2andtheprobabilityofjk1)]TJ /F6 11.955 Tf 14.2 2.66 Td[(^k1j>^k1)]TJ /F3 11.955 Tf 11.95 0 Td[(lisnogreaterthan1.OursimulationparametersaresetbasedonthetypicalsettingoftheEPCglobalGen-2standard[ 14 ].Anytwoconsecutivetransmissions(fromareadertotagsorfromatagtothereader)areseparatedbyawaitingtimeof266.4s.Accordingtothespecication,thetransmissionratefromatagtothereaderisthesameasthetransmissionratefromthereadertoatag.Theratefromatagtothereaderis40.97Kb=sec;ittakes24.4sforatagtotransmitonebit.Thelengthofaslotiscalculatedasthesumofawaitingtimeandthetimeforthetagtotransmitacertainnumberofbits.ThetypeofslotsusedbyTBC,ARTandZOE,denotedasTshort,containsonlyonebit.Itslengthis290.8s.ThetypeofslotsusedbyGTneedstodetectcollisionandcontains10bits.TheirslotlengthisTlong=470.5s.Comparingwiththetotalamountoftimeusedbyalltagstotransmittothereader,thetimeusedbythereadertobroadcastinformationtothetagsisnegligibleinallthreeprotocols.ForTBC(p1),weapproximatelycompute( 5 )withveloopsforp1from0to1atstepsof0.001,wfromzeroto50,f1from0to10n1,m1from0to1000,andTfrom 100

PAGE 101

ltoh.Therangesaresetbasedonourempiricalinvestigationwhichshowsthattheoptimalparameterspersistentlyfallwithintheseranges.Wendthebestvalueofwf1(togetherwiththecorrespondingsystemparameters)intheseranges.Oncewf1isdetermined,theexecutiontimeisknown,whichisTshortwf1plusthetimeofestimatingthevalueofn1usingZOE;notethateventhroughn1ispre-determinedforeachsimulation,weassumethatthereaderdoesnotknowthisvaluebeforehandanditrunsZOEoncetoestimaten1with5%errorat95%condencelevel.TBC(100%)worksjustlikeTBC(p1),exceptthatitisoperatedunderthesamplingprobabilityp1=1.GTwillalsocomputeitsoptimalsystemparameters,includingthetimeframesizef1,thenumberRofrounds,andthenumberofshufedgroupsW.TheexecutiontimerequiredbyGTisTlongf1WR.TheparametersofARTandZOEaresetbasedontheiroriginalpapers.Inoursimulations,n1=500,000,therangeofgroupsizesis(0,500],h=250,andthevalueoflvaries.Thereare2,000groups.Thenumberxofabove-thresholdgroups(whosesizesaregreaterthanh)mayvaryinsomesimulations,butitsdefaultvalueis1,000.Besidesrandomlychoosingthesizeofeachabove-thresholdgroupfrom[250,500]andthatofeachbelow-thresholdgroupfrom[1,249]inonesimulation,ourdefaultwayofdetermininggroupsizesisgivenasfollows:Werstrandomlychoosethesizesfortheabove-thresholdgroupsfrom[250,500].Afterthat,wedistributetheremainingMtagsintothebelow-thresholdgroups.Fortherstbelow-thresholdgroup,wegeneratearandomnumberbetween1andminf249,M 2000)]TJ /F7 7.97 Tf 6.58 0 Td[(xgtobeitspopulation,whichisdenotedass1.Forthesecondbelow-thresholdgroup,weselectarandomvaluebetween1andminf249,M)]TJ /F7 7.97 Tf 6.58 0 Td[(s1 1999)]TJ /F7 7.97 Tf 6.58 0 Td[(xgasitspopulation,whichisdenotedass2.Similarly,weassignarandomvaluebetween1andminf249,M)]TJ /F7 7.97 Tf 6.59 0 Td[(s1)]TJ /F7 7.97 Tf 6.58 0 Td[(s2 1998)]TJ /F7 7.97 Tf 6.59 0 Td[(xgasthepopulationforthethirdbelow-thresholdgroup.Thisprocessisrepeatedforallremainingbelow-thresholdgroups.Iftherearestilltagsleftunassigned,weassignthemarbitrarilytobelow-thresholdgroupsaslongastheirsizesarebelow250. 101

PAGE 102

5.4.2ExecutionTimeRequiredwithRespectto1,1andl=hWecompareTBC(100%),TBC(p1),GT,ARTandZOEintermsofexecutiontime.Tables 5-2 5-5 showoursimulationresultsunderdifferentvaluesofl=h,1and1.Table 5-2 showstheexecutiontimerequiredwhen1=99.9%and1=0.1%.Fromthetable,wecanseethatTBC(100%)hasamuchsmallerexecutiontimethanGT,ARTandZOE.Forexample,GTtakes3304300whenl=0.5h,whichisabouttripleofthetimetakenbyTBC(100%).Whensamplingisintroduced,theperformanceofTBCimproves,i.e.,TBC(p1)takeslesstimetoclassifytheabove-thresholdgroupsunderthesamesetting-only27.2%ofthetimetakenbyGT.ARTandZOEconsumeanorderofmagnitudeormoretimethanTBC(p1).Whenl=hbecomeslarger,TBC(100%),TBC(p1),GT,ARTandZOEneedmoretimetoclassifytheabove-thresholdgroups.Thisisbecausealargerratioofl=hmeansahigheraccuracyrequirementforclassication.TheperformancegainbyTBC(p1)overGT,ART,ZOEandTBC(100%)shrinksasl=hincreases,butremainssignicant.Forexample,whenl=0.7h,theexecutiontimerequiredbyTBC(p1)is31.2%ofthetimebyGT.Whenl=0.9h,thetimebyTBC(p1)is34.1%ofthatbyGT.GTusesasimple,fastthresholdcheckingschemetoprobabilisticallyidentifypopulousgroupswithsizelargerthanathreshold.However,itincursalargevarianceinitsestimatedresult.Tosatisfyahighaccuracyrequirement,alargenumberofexecutionsarerequired,whichlengthensexecutiontime.Inaddition,GThastoidentifywhetheraslotisempty,singletonorcollision,resultinginlongerslots.TBC(100%)estimatesallgroupsizestogetherandshareslotsamongallgroups.Inaddition,itonlyneedstoknowwhethereachslotisemptyornot.Hence,itsexecutiontimeisshorter.SamplingisturnedoninTBC(p1),whichrequiresonlyafractionoftagstoparticipateinaprotocolexecution.Whenthesamplingprobabilityp1ischosentobeasmallvalue,thenumberofparticipatingtagsarelargelyreduced,whichinturnreducestheframesize 102

PAGE 103

Figure5-2. ExecutiontimewithrespecttothenumberofgroupsSwhicharesupposedtobereportedwhen1=99%,1=1%andl=0.7h.Thetotalnumberoftagsn1isxedtobe500,000ateachpoint. Figure5-3. ExecutiontimewithrespecttothenumberofgroupsSwhicharesupposedtobereportedwhen1=99%,1=1%andl=0.7h.Thetotalnumberoftagsn1forallthegroupsincreasesalongwithS. foreachpolling.ItisnotefcienttoinvokeARTandZOEtoestimatethesizeofeachgrouponeatatime.Tables 5-3 and 5-5 comparetheexecutiontimesoftheveprotocolswhen1=99%,1=1%,1=95%,1=5%,and1=90%,1=10%,respectively.ThesethreetablesshowthatTBC(p1)outperformsotherprotocolsunderdifferentparametersettings.WhencomparingwithTable 5-2 ,weseethatgiventhesamevaluesofhandl,theexecutiontimesofallprotocolsarereducedwhen1decreasesor1increases.ButtheperformancegainbyTBC(p1)remainssignicant. 5.4.3FPRandFNRwithRespectto1,1andl=hWecallagroupwhosesizeisnomorethanl(nolessthanh)asabelow-l(above-h)group.Thefalsepositiveratio(FPR)isdenedasthefractionofbelow-lgroupsthataremistakenlyreported.Thefalsenegativeratio(FNR)isdenedas 103

PAGE 104

Table5-2. Estimationtimecomparisonwhen1=99.9%and1=0.1% EstimationTimeinminutes(0),seconds(00)TBC(100%)TBC(p1)GTARTZOE l=0.1h702300404300200230057054004801000l=0.3h805900605600250570068018005701900l=0.5h1001500901900330430081050006901300l=0.7h1402300120120042012009605008003000l=0.9h20034001705700610390011401900950100 Table5-3. Estimationtimecomparisonwhen1=99%and1=1% EstimationTimeinminutes(0),seconds(00)TBC(100%)TBC(p1)GTARTZOE l=0.1h101400390030490039025002601400l=0.3h104200570040260046031002800500l=0.5h30050010440060150061050003505600l=0.7h40510020520090120084034004701500l=0.9h6021004005001105800100024005902300 Table5-4. Estimationtimecomparisonwhen1=95%and1=5% EstimationTimeinminutes(0),seconds(00)TBC(100%)TBC(p1)GTARTZOE l=0.1h44001800105500903100704700l=0.3h106002900209001101200802300l=0.5h1047001001002057001201000902600l=0.7h20390010430040110014021001101400l=0.9h4040020560060360017011001403200 Table5-5. Estimationtimecomparisonwhen1=90%and1=10% EstimationTimeinminutes(0),seconds(00)TBC(100%)TBC(p1)GTARTZOE l=0.1h38001500103300503700503800l=0.3h52002000105100601900504700l=0.5h1017004800201800705600603300l=0.7h1056001014003060080430080500l=0.9h20500020250040550010024001001100 104

PAGE 105

Table5-6. FalsenegativeRatioandFalsePositiveRatiowhen1=99%and1=1% FNRFPRTBC(100%)TBC(p1)GTTBC(100%)TBC(p1)GT l=0.1h0.00840.00950.00980.00750.00870.0086l=0.3h0.00840.00950.00980.00750.00890.0090l=0.5h0.00840.00950.00980.00790.00890.0090l=0.7h0.00840.00950.00980.00810.00900.0090l=0.9h0.00840.00950.00980.00890.00910.0093 Table5-7. FalsenegativeRatioandFalsePositiveRatiowhen1=95%and1=5% FNRFPRTBC(100%)TBC(p1)GTTBC(100%)TBC(p1)GT l=0.1h0.0370.0420.0460.040.0390.041l=0.3h0.0370.0420.0460.0410.0420.042l=0.5h0.0370.0420.0460.0430.0430.042l=0.7h0.0370.0420.0460.0430.0460.042l=0.9h0.0370.0420.0460.0440.0470.045 Figure5-4. Executiontimewithrespecttothenumberofgroupswhen1=99%,1=1%,h=250,l=0.5h,andn1increaseswiththenumberofgroups. Figure5-5. Executiontimewithrespecttothenumberofgroupswhen1=99%,1=1%,h=250,l=0.5h,andn1isxedat500000. 105

PAGE 106

Figure5-6. Zoom-inofFigure 5-5 forTBCandGT. thefractionofabove-hgroupsthatarenotreported.Tables 5-6 5-7 presentoursimulationresultsofFNRandFPRunderdifferentvaluesof1,1andl=h.ForTable 5-6 ,1=99%and1=1%.TheFNRvaluesofTBC(100%),TBC(p1)andGTareconsistentlysmallerthan1)]TJ /F4 11.955 Tf 12.41 0 Td[(1andtheirFPRvaluesareconsistentlysmallerthan1,whichmeansthattheseprotocolsmeettheperformanceobjectivesin(1).Inaddition,weobservethatourprotocolhassmallerFNRandFPRthanGT.Theresultsfor1=95%and1=5%areTable 5-7 ,wherethevaluesofFPRandFNRalsomeettheobjectives. 5.4.4ExecutionTimeRequiredwithRespecttoAbove-ThresholdGroupsInthepreviouscomparison,thenumberofabove-thresholdgroupsissetatthedefaultvalue1,000.WefurthercompareTBC(100%),TBC(p1)andGTbyvaryingthenumberofabove-thresholdgroups,denotedasS.Let1=99%and1=1%.InFig. 5-2 ,wekeepn1=500,000andvarythethenumberofabove-thresholdgroupsfrom250to1,250.Fromthegure,theexecutiontimeofGTislinearinS,butTBC(100%)andTBC(p1)aredifferent.NotonlydotheyoutperformGT,butalsotheirexecutiontimesarebothinsensitivetoS.Aslongasthetotalnumberoftagsinthesystemisthesame,theirexecutiontimescanbeapproximatelyviewedasconstantsevenwhenthenumberofgroupsisdifferent.Suchanobservationagreeswith( 5 ),whichdoesnotincludeSinitsformulation.Furthermore,thankstosampling,ittakesTBC(p1)lesstimetoclassifytheabove-thresholdgroupsthanTBC(100%),foranyvalueofS. 106

PAGE 107

InFig. 5-3 ,wexthenumberofgroupsto2,000,whileallowingthetotalnumberoftagstochange.Eachbelow-thresholdgrouptakesarandompopulationintherangeof(0,250)andeachabove-thresholdgrouptakesarandompopulationintherangeof[250,500].Fromthegure,weobservethattheexecutiontimesofTBC(100%),TBC(p1),andGTareapproximatelyproportionaltothenumberofabove-thresholdgroups.However,thelinesofTBC(p1)andTBC(100%)havesomewhatsmallerslopesthanthelineofGT. 5.4.5ExecutionTimeRequiredwithRespecttoTheNumberofGroupsWeletthenumberofgroupsinthesystemtoincreasefrom250to2000,withtheirgroupsizesrandomlyselectedfrom(0,500].Fig. 5-4 showstheexecutiontimesoftheveprotocolswithrespecttothenumberofgroups.Alltheprotocolstakelongertimetoclassifymoregroups.However,theexecutiontimesofTBC(p1)andTBC(100%)increasewithsmallerslopes,andthereforetheyaremorescalablethanZOE,ARTandGT.Forexample,whenthereare4,000groups,thetimeforARTis27,974seconds,thetimeforZOEis15,582seconds,thetimeforGTis753seconds,andthetimeforTBC(100%)is451seconds,andthetimeforTBC(p1)is315seconds.NextweshowaninterestingpropertyofTBCwhenwexthetotalnumberoftagsinthesystemat500,000butvarythenumberofgroups.Fig. 5-5 showstheexecutiontimesoftheveprotocolswithrespecttothenumberofgroups.ThetimesofARTandZOEstillgrowroughlylinearlywiththenumberofgroups.WereplotthecurvesofGT,TBC(p1)andTBC(100%)inFig. 5-6 forabetterview.TheexecutiontimeofGTgrowswithrespecttothenumberofgroups.ButthetimesofTBC(p1)andTBC(100%)stayat,insensitivetothenumberofgroupswhenthetotalnumberoftagsisunchanged.Thereasonissimple:Eachgroupisassignedacertainnumberofslots,andeachtaginothergroupsmaycausenoiseinoneofthoseslotsduetosharing.ForTBC,theoverallnoiseonlydependsonthetotalnumberoftagsinothergroups,notthenumberofothergroups. 107

PAGE 108

5.5SummaryThisdissertationproposesanewsolutionformultigroupthreshold-basedclassicationinalargeRFIDsystem.Whilemuchofthepriorworkfocusesonestimatingthetotalnumberoftagsinasystem,itisinefcienttoapplythosesolutionstosequentiallyestimatingthesizeofeachtaggroupandseeifitisaboveathreshold.Inthisdissertation,weproposeanewprotocolbasedonlogicalbitmapsthatallowthesizesofallgroupstobeestimatedtogetherforclassication.Slotsharingisexploitedtoreducetheexecutiontime.Furthermore,samplingisintroducedtoreducethenumberofparticipatingtags(andthuscollision),whichinturncutsdowntheexecutiontime.Reducingthenumberofparticipatingtagsalsosavestheoverallenergyexpenditurebythetagsifbattery-poweredactivetagsareused.Ournewprotocolisabletoperformmultigroupclassicationwithanypre-setaccuracy.Weevaluatetheproposedsolutionanddemonstratethatitcomparesfavorablywiththebestexistingwork.Wealsopresentthemethodofcomputingtheoptimalsystemparameters. 108

PAGE 109

CHAPTER6IMPLEMENTATION 6.1SystemSetupWeimplementaprototypesystemtovalidateTRP-asimplemissing-tagdetectionprotocol,andasimplemulti-groupclassicationprotocol,usingprogrammableMOOtagsbasedontheWISP4.1hardwareandrmware,asshowninFig 6-1 .TheMOOtagmainlyconsistsofanRFIDcircuitryandanultra-lowpower16-bitMSP430microcontroller,whichareusedtoharvestpowerandbackscatterradiosignals.Inaddition,weusedanImpinjSpeedwayRFIDreadertointerrogatewiththeMOO,withthereader'stransmitpowersetto+30dBmanda6-dBicircularlypolarizedpatchantennaattached.CurrentMOOrmwarehaspartiallyimplementedtheEPCglobalGen-2standardoperatinginthe902-928MHzindustrial-scientic-medical(ISM)band[ 14 ].WeextendtheEPCglobalC1G2protocolwiththefunctionalityofTRPandcardinalityestimation.Actually,theimplementationoftheTRPprotocolandcardinalityestimationonlyrequiresaslightextensiontotheEPCglobalC1G2protocol.WendthatanAloha-basedanti-collisionpollingschemehasbeenstandardizedbyEPCglobalwherethereaderbeginseachinterrogationroundbyinformingallthetagsabouttheframesize.Eachtagthenchoosesatimeslotatrandomandtransmitsonlywithinthattimeslot.Morespecically,inEPCglobalC1G2standard[ 14 ],areaderinitiateseachcommunicationroundwithtags.Thereadertransmitsanoperationcode(e.g.,Query,Write,Select,ACKetc.)indicatingtheexpectedoperationoftags,thebackscatterbitrate,andtagencodingschemes(e.g.,FM0orMiller).Forexample,thereadermayinitiatecommunicationbysendingtagsaQuerycommand,whichincludesaeldthatsetsthenumberf1ofslotsintheround.Inaddition,thiscommandalsoinvolvesotherparametersthatcanbeusedtonegotiatewiththetagsaboutthelengthofeachslotandthewaitingtimebetweenconsecutiveslots.UponreceivingtheQuery,eachtagshouldpickarandomvalueintherange[0,2f1),andloadthisvalueintoitsslot 109

PAGE 110

Figure6-1. RFIDSystem Figure6-2. Timeframelength=100slots. counter.ItdecrementstheslotcountereachtimewhenreceivingaQueryRepfromthereader.Ifitsslotcounterisdecreasedtozero,itrepliestothereader;otherwisethetagshallremainsilent.TheFig. 6-2 andFig. 6-3 showanexampleofrepliesfromtagsusinganti-collisionpollingschemewith20tags.Thehorizontalaxisistheslotsequencenumber,andtheverticalaxisisthemagnitudeofreplies.Wenormalizethetransmissionsignaloftags,anduse'1'torepresentthesignalstrengthwhenthereisanyreplyinaslot.FromFig. 6-2 ,wecanseethe20tagsreplyin20slotsoutof100,resultinginnocollision.Whentheframelengthisonly30,thecollisionhappens,whichcanbeobservedfromFig. 6-3 wherethereare5collisionslots. 6.2ASimpleMissing-tagDetectionPorotocolImplementationWeimplementtheTRPprotocolbyfollowingtheconventionalreader-initiatedapproach.WerstaddtheDetectcommandintothecommandsetofthestandard.Todetectthemissingtagevent,thereaderinitiatescountingprocedurebysendingaDetectcommandalongwithotherparameters(r,andencodingscheme,etc).Inthecasethat 110

PAGE 111

Figure6-3. Timeframelength=30slots. Figure6-4. BeforevetagsareMssing. theoperationcodeisDetect,eachtagcomputesahashvaluewiththerandomnumberanditstagID,thentransmitsashortresponseaccordingtotheencodingscheme.WedidtheexperimentwithtwentyMOOtags,whichisshowninFig. 6-4 andFig. 6-5 .Theformergureshowstheslotstatusinthecasethatalltagsareinthesystem,whilethelatteronepresentstheresultswhenvetagsaremissing.FromFig. 6-4 ,wecanseethereaderreceivesrepliesfromalltwentytags,whichcanbeidentiedbytwentysingletonslots.InFig. 6-5 ,thereadergetsonly15responsebecause5tagsaremissing,whichcanbeobservedfromthe15non-emptyslots. Figure6-5. AftervetagsareMissing. 111

PAGE 112

Figure6-6. Thepopulationestimationoffourgroups. 6.3ASimpleMulti-groupClassicationPorotocolImplementationInthissection,wewanttoshowtheresultsofoursimplemulti-groupclassicationprotocolimplementation.WestilluseaMOOtosimulatevevirtualMOOs,andthevirtualMOOssimulatedonthesamephysicalMOObelongstothesamegroup(withthesamegroupID).Foreachgroup,anindividualtimeframeisusedtoestimateitstagpopulationbycountingthenumberofnon-emptyslotsintheframe.TheresultisshowninFig. 6-6 ,wherethehorizontalaxisisthegroupsequencenumber,andtheverticalaxisistheestimatedgrouppopulation.Fromthegure,wecanseethattheestimatednumberoftagsintherstgroupis4,lessthantheactualvalue.Thereasonisthattwotagsreplytothereaderatthesameslot.Thesecondgroupisestimatedtohave3tags,thethirdgrouphasfourtagsandthefourthonehasvetags.Ifwesetupthethresholdvaluetobe3,allthegorupexcepthesecondonewillbereportedasabove-thresholdgroups. 6.4What'stobeexpected?Fromtheaboveresults,wecanseeitispossibletoimplementprotocolstodetectmissing-tageventorclassifytaggroupsinthescenarioofsmall-scalesystems.TheimplementationactuallycanbescaledtoworkinalargeRFIDsystem,justbyexpandingthetimeframetohavemoretimeslots.Inthiscase,eachtagworksinthesamewaytoselectaslotrandomlyandreply.Basedonthestatusoftimeslots,themissingtagcan 112

PAGE 113

bedetected,orthenumberoftagsineachtaggroupcanbeestimatedandappliedtoclassifydifferenttaggroups. 113

PAGE 114

CHAPTER7CONCLUSIONANDPROPOSALFORFUTUREWORK 7.1ConclusionsInthisdissertation,werstproposesanewprotocoldesignthatintegratesenergyefciencyandtimeefciencyformissing-tagdetection.ItusesmultiplehashseedstoprovidemultipledegreesoffreedomfortheRFIDreadertoassigntagstosingletonslots,duringwhichthetagsannouncetheirpresenceintheprocessofmissing-tagdetection.Wealsopresentthisnewprotocolwithreliablechannels.Theresultisamulti-foldcutinbothenergycostandexecutiontime.Inaddition,weextendtheprotocoltoconsidertwocategoriesofchannelerrorsinducedbynoise/interferenceintheenvironment.Theinvolvingofchannelerrorswillmaketheenergy/timegainsslightlyreduced,butremainsignicantcomparingtoEMDandTRP.Wealsorevealafundamentalenergy-timetradeoffintheprotocoldesign.Thistradeoffgivesexibilityinperformancetuningwhentheprotocolisappliedinpracticalenvironment.Wethenfocusondesigninganewprotocolthatcansolvethemultigroupthreshold-basedclassicationprobleminalargeRFIDsystem.Whilemuchofthepriorworkfocusesonestimatingthetotalnumberoftagsinasystem,itisinefcienttoapplythosesolutionstosequentiallyestimatingthesizeofeachtaggroupandseeifitisaboveathreshold.Inthisdissertation,weproposeanewprotocolbasedonlogicalbitmapsthatallowthesizesofallgroupstobeestimatedtogetherforclassication.Slotsharingisexploitedtoreducetheexecutiontime.Furthermore,samplingisintroducedtoreducethenumberofparticipatingtags(andthuscollision),whichinturncutsdowntheexecutiontime.Reducingthenumberofparticipatingtagsalsosavestheoverallenergyexpenditurebythetagsifbattery-poweredactivetagsareused.Ournewprotocolisabletoperformmultigroupclassicationwithanypre-setaccuracy.Weevaluatetheproposedsolutionanddemonstratethatitcomparesfavorablywiththebestexistingwork.Wealsopresentthemethodofcomputingtheoptimalsystemparameters. 114

PAGE 115

Intheend,weimplementasimplemissing-tagdetectionprotocolandmulti-groupthreshold-basedclassicationinthephysicaldevices-RFIDMOOtagsandspeedwayRFIDreaders.Theresultsshowthatthereadercancooperatewithtagstodetectmissing-tageventbasedonthesimplealoha-basedprotocol.Futuremore,byestimatingthenumberoftagsineachgroup,thereaderisalsoabletoclassifytaggroupsinasmall-scalescenario. 7.2FutureWorkAscanbeseen,theproposedprotocolinchapter4canbeusedtosolvetheproblemofRFIDmultigroupthreshold-basedclassication.Inthisdissertation,wepresenthowtoclassifyRFIDtaggroupsbasedondynamicslotingsharing,andthenshowanalyticallyhowtocomputeoptimalsystemparametersthatminimizetheprotocolexecutiontimeundertheconstraintoftherequirement.Wemayextendthisproblemfromtwodifferentangles:First,consideralargewarehousewithlotsofproducts,whereareaderisnotabletocoverthewholewarehousetoclassifyalltheproducts.Inthisscenario,wemayneedtoinstallmultiplereaderssothateachreaderisresponsiblefortheproductswithinitsinterrogatorregion.However,thecoverageofnearbyreadersmayoverlap,andthereforetagsinagroupmayappearintherangesofdifferentreaders.WhenthereadersexecuteTBC(p)orTBC(100%)separately,thegroupmaybeclassiedintodifferentcategoriesbydifferentreaders,relyingonthenumberoftagscovered.Forexample,agroupisintheinterrogatorrangeoftworeaders,AandB.ThereaderAismonitoring300tagsofthisgroup,whilethereaderBhas100tags.Supposethethresholdvalueis200.Classiedbyourprotocol,theportioncoveredbyAisidentiedasanabove-thresholdgroup,andtheotherportionisanormalgroup.Hence,itisnecessarytocomeupwithabettersolutionthatcanaddresssuchdisruptionclassicationresultsgeneratedbybyTBC(p)orTBC(100%),duetotheintersectioncoveredbytwoormorenearbyreaders. 115

PAGE 116

Next,inpractice,differentgroupsmayneedtobeclassiedbasedondifferentthresholds.Forexample,inawarehousethatstoresshoes,computersandotherproducts,thethresholdsforshoesandcomputersmaybedifferentbecausetheirexpectedinventorylevelsmaynotbethesame.Theproblembecomesclassifyingthegroupsintodifferentrangesbasedontheirsizes.Whentherearemultiplethresholds,werstplacetaggroupsinsubsets,eachofwhichcorrespondstoathreshold.Wethenperformsingle-thresholdclassicationwithineachsubset.Todoso,thereaderbroadcaststhegroupIDsinthesubset,andtheclassicationisperformedamongthetagsthatcarryoneofthoseIDs.Thissolutioniseasytoimplementbasedonourproposedprotocolratherthanoptimizedbecauseweseetheperformancedegradeswhentherearealargenumberofthresholds.Thereasonisthattheexecutiontimeisinsensitivetothethenumberofgroups.Whenthewholesystemisdividedintoalargenumberofsubsetsbasedondifferentthresholds,thetotalexecutiontimecanbecalculatedbytheproductofthenumberofthresholdsandtheexecutiontimetoclassiedgroupsinasubset.Forexample,wesupposeasubsetcontainsgroupsbetween1500and2500.ObservedfromFig. 5-5 ,ittakesTBC(p)180secondstoclassifygroupsinasubsetwhen=99%,=1%,h=250,l=0.5h,andnisxedat500,000.Ifwehave20thresholdvalues,thetotalexecutiontimeis20180=3,600seconds,whichisanintoleranttimeduration.Hence,theRFIDmultigroupthreshold-basedclassicationwithdifferentthresholdsisstillunderexploration,andthefutureworkshouldfocusonndingabettersolutionthatcancutdownthetimerequired. 116

PAGE 117

REFERENCES [1] Diracdeltafunction.(2000).URL http://en.wikipedia.org/wiki/Dirac_delta_function [2] InformationtechnologyautomaticidenticationanddatacapturetechniquesCradiofrequencyidenticationforitemmanagementairinterface-part6:parametersforairinterfacecommunicationsat860-960MHz.FinalDraftIn-ternationalStandardISO18000-6(2003). [3] Barcode.Barcode.(2014).URL http://en.wikipedia.org/wiki/Barcode [4] Bhandari,N.,Sahoo,A.,andIyer,S.IntelligentQueryTree(IQT)ProtocoltoImproveRFIDTagReadEfciency.Proc.ofIEEEInternationalConferenceonInformationTechnology(2006). [5] Bustillo,M.Wal-martRadioTagstoTrackClothing.(2010).URL http://online.wsj.com/article/SB10001424052748704421304575383213061198090.html [6] Buzzle.com.PassiveVs.ActiveRFIDTags.(2011).URL http://www.buzzle.com/articles/passive-vs-active-rfid-tags.html [7] Capetenakis,J.I.TreeAlgorithmsforPacketBroadcastChannels.IEEETransactionsonInformationTheory25(1979).5:505. [8] Cha,J.andKim,J.NovelAnti-collisionAlgorithmsforFastObjectIdenticationinRFIDSystem.Proc.IEEEICPADS(2005). [9] Cha,J.R.andKim,J.H.DynamicFramedSlottedALOHAAlgorithmsUsingFastTagEstimationMethodforRFIDSystems.Proc.ofIEEEConsumerCommunica-tionsandNetworkingConference(2006). [10] Chen,S.andNahrstedt,K.MaxminFairRoutinginConnection-OrientedNetworks.Proc.ofEuro-ParallelandDistributedSystemsConference(1998):163. [11] Chen,S.andShavitt,Y.SoMR:AscalabledistributedQoSmulticastroutingprotocol.JournalofParallelandDistributedComputing68(2008).2:137. [12] Choi,H.,Cha,J.,andKim,J.FastWirelessAnti-collisionAlgorithminUbiquitousIDSystem.Proc.IEEEVTC(2004). [13] Cornaglia,B.andSpini,M.Letter:Newstatisticalmodelforbursterrordistribution.EuropeanTransactionsonTelecommunications7(1996).3:267. 117

PAGE 118

[14] EPCglobal.EPCRadio-FrequencyIdentityProtocolsClass-1Generation-2UHFRFIDProtocolforCommunicationsat860MHz-960MHzVersion1.0.9.(2005).URL http://www.epcglobalinc.org/standards/uhfc1g2/uhfc1g2_1_0_9-standard-20050126.pdf [15] Firner,B.,Jadhav,P.,Zhang,Y.,Howard,R.,Trappe,W.,andFenson,E.TowardsContinuousAssetTracking:Low-PowerCommunicationandFail-SafePresenceAssurance.Proc.ofthe6thAnnualIEEECommunicationsSocietyConferenceonSensor,MeshandAdHocCommunicationsandNetworks(2009). [16] Han,H.,Sheng,B.,Tan,ChiuC.,Li,Q.,Mao,W.,andLu,S.CountingRFIDTagsEfcientlyandAnonymously.Proc.IEEEINFOCOM(2010). [17] Hush,D.andWood,C.AnalysisofTreeAlgorithmforRFIDArbitration.Proc.ofIEEEISIT(1998). [18] Jacobsen,R.M.,Nielsen,K.F.,Popovski,P.,andLarsen,T.ReliableIdenticationofRFIDTagsusingMultipleIndependentReaderSessions.Proc.ofIEEERFID(2009). [19] Jian,Y.,Chen,S.,Zhang,Z.,andZhang,L.Anovelschemeforprotectingreceiver'slocationprivacyinwirelesssensornetworks.IEEETransactionsonWirelessCommunications7(2008).10:3769. [20] Klair,D.,Chin,K.,andRaad,R.OntheEnergyConsumptionofPureandSlottedAlohabasedRFIDAnti-CollisionProtocols.ComputerCommunications(2008). [21] Kodialam,M.andNandagopal,T.FastandReliableEstimationSchemesinRFIDSystems.Proc.ofACMMobiCom(2006). [22] Kodialam,M.,Nandagopal,T.,andLau,W.AnonymousTrackingusingRFIDtags.Proc.ofIEEEINFOCOM(2007). [23] Lee,S.R.,Joo,S.D.,andLee,C.W.AnEnhancedDynamicFramedSlottedALOHAAlgorithmforRFIDTagIdentication.Proc.ofIEEEInternationalConfer-enceonMobileandUbiquitousSystems:NetworksandServices(2005). [24] Li,T.,Chen,S.,andLing,Y.IdentifyingtheMissingTagsinaLargeRFIDSystem.Proc.ofACMMobihoc(2010). [25] Li,T.,Luo,W.,Mo,Z.,andChen,S.Privacy-preservingRFIDAuthenticationbasedonCryptographicalEncoding.Proc.ofIEEEINFOCOM(2012):21742182. [26] Li,T.,Wu,S.,Chen,S.,andYang,M.EnergyEfcientAlgorithmsfortheRFIDEstimationProblem.Proc.ofIEEEINFOCOM(2010). [27] .GeneralizedEnergy-EfcientAlgorithmsfortheRFIDEstimationProblem.IEEE/ACMTransactionsonNetworking20(2012):19781990. 118

PAGE 119

[28] Liu,X.,Li,K.,Shen,Y.,Min,G.,Xiao,B.,andQu,W.AFastApproachtoUnknownTagIdenticationinLargeScaleRFIDSystems.Proc.ofICCCN(2013). [29] Luo,W.,Chen,S.,Li,T.,andChen,S.EfcientMissingTagDetectioninRFIDSystems.Proc.ofIEEEINFOCOM,mini-conference(2011). [30] Luo,W.,Qiao,Y.,andChen,S.AnEfcientProtocolforRFIDMultigroupThreshold-basedClassication.Proc.ofIEEEINFOCOM(2013). [31] M.Chen,Z.Mo,S.ChenW.LuoandFang,Y.AnEfcientTagSearchProtocolinLarge-ScaleRFIDSystems.Proc.ofIEEEINFOCOM(2013). [32] Myung,J.andLee,W.Anadaptivememorylesstaganti-collisionprotocolforRFIDnetworks.Proc.IEEEICC(2005). [33] .AdaptiveSplittingProtocolsforRFIDTagCollisionArbitration.Proc.ofACMMOBIHOC(2006). [34] Namboodiri,V.andGao,L.Energy-AwareTagAnti-CollisionProtocolsforRFIDSystems.Proc.ofIEEEPerCom(2007). [35] Ni,L.M.,Liu,Y.,Lau,Y.C.,andPatil,A.LANDMARC:IndoorLocationSensingUsingActiveRFID.Proc.ofACMWirelessNetworks10(2004).6. [36] Popovski,P.,Nielsen,K.F.,Jacobsen,R.M.,andLarsen,T.RobustStatisticalMethodsforDetectionofMissingRFIDTags.IEEECommunicationsMagazine,WirelessCommunicationsSeries(2010). [37] Qian,C.,Ngan,H.,andLiu,Y.CardinalityEstimationforLarge-scaleRFIDSystems.Proc.ofIEEEPerCom(2008). [38] Qiao,Y.,Chen,S.,andLi,T.RFIDasanInfrastructure.SpringerISBN:978-1-4614-5229-4(2012). [39] Qiao,Yan,Chen,Shigang,Li,Tao,andChen,Shiping.Energy-efcientPollingProtocolsinRFIDSystems.Proc.ofACMMobiHoc(2011). [40] S.Chen,M.ZhangandXiao,B.EfcientInformationCollectionProtocolsforSensor-augmentedRFIDNetworks.Proc.ofIEEEINFOCOM(2011). [41] Sarangan,V.,Devarapalli,M.R.,andRadhakrishnan,S.AFrameworkforFastRFIDTagReadinginStaticandMobileEnvironments.InternationalJournalofComputerandTelecommunicationsNetworking52(2008).5:1058. [42] Sato,Y.,Igarashi,Y.,Mitsugi,J.,Nakamura,O.,andMurai,J.CountingRFIDTagsEfcientlyandAnonymously.Proc.ofIEEERFID(2012). [43] Shahzad,M.andLiu,A.EveryBitCounts-FastandScalableRFIDEstimation.Proc.ofACMMOBICOM(2012). 119

PAGE 120

[44] Sheng,B.,Li,Q.,andMao,W.EfcientContinuousScanninginRFIDSystems.Proc.ofIEEEINFOCOM(2010). [45] Sheng,B.,Tan,ChiuC.,Li,Q.,andMao,W.FindingPopularCategoriesforRFIDTags.Proc.ofACMMOBIHOC(2008). [46] Tan,C.,Sheng,B.,andLi,Q.HowtoMonitorforMissingRFIDTags.Proc.ofIEEEICDCS(2008). [47] Telatar,I.E.andGallager,R.G.CombiningQueuingTheoryandInformationTheoryforMultiaccess.IEEEJournalonSelectedAreasCommunication13(1995).6:963. [48] Vogt,H.EfcientObjectIdenticationwithPassiveRFIDTags.Proc.ofIEEEPerCom(2002). [49] .EfcientObjectIdenticationwithPassiveRFIDTags.Proc.ofIEEEPerCom(2002). [50] Wolverton,J.Wal-MarttoEmbedRFIDTagsinClothingBeginningAugust1.(2010).URL http://www.thenewamerican.com/index.php/tech-mainmenu-30/computers/4157-wal-mart-to-embed-rfid-tags-in-clothing-beginning-august-1 [51] Xiao,Q.,Xiao,B.,andChen,S.DifferentialEstimationinDynamicRFIDSystems.Proc.ofIEEEINFOCOM,mini-confernece(2013). [52] Yao,Q.,Qi,Y.,Han,J.,Zhao,J.,Li,X.,andLiu,Y.RandomizingRFIDPrivateAuthentication.Proc.ofIEEEPERCOM(2009). [53] Yue,H.,Zhang,C.,Pan,M.,Fang,Y.,andChen,S.ATime-efcientInformationCollectionProtocolforLarge-scaleRFIDSystems.Proc.ofIEEEINFOCOM(2012). [54] Zhai,J.andWang,G.N.AnAnti-collisionAlgorithmusingTwo-functionedEstimationforRFIDTags.Proc.ofICCSA(2005). [55] Zhang,M.,Li,T.,Chen,S.,andLi,B.UsingAnalogNetworkCodingtoImprovetheRFIDReadingThroughput.Proc.ofIEEEICDCS(2010):547556. [56] Zhang,R.,Liu,Y.,Zhang,Y.,andSun,J.FastIdenticationoftheMissingTagsinaLargeRFIDSystem.Proc.ofthe8thAnnualIEEECommunicationsSocietyConferenceonSensor,MeshandAdHocCommunicationsandNetworks(2011). [57] Zhen,B.,Kobayashi,M.,andShimizu,M.FramedALOHAforMultipleRFIDObjectIdentication.IEICETransactionsonCommunications(2005). [58] Zheng,Y.andLi,M.FastTagSearchingProtocolforLarge-ScaleRFIDSystems.Proc.ofIEEEICNP(2011). 120

PAGE 121

[59] .ZOE:FastCardinalityEstimationforLarge-ScaleRFIDSystems.Proc.ofIEEEINFOCOM(2013). [60] Zhou,F.,Chen,C.,Jin,D.,Huang,C.,andMin,H.EvaluatingandOptimizingPowerConsumptionofAnti-collisionProtocolsforApplicationsinRFIDSystems.Proc.ofACMInternationalSymposiumonLowPowerElectronicsandDesign(ISLPED)(2004). 121

PAGE 122

BIOGRAPHICALSKETCH WenLuoreceivedhisB.S.degreeincomputerscienceandtechnologyfromtheUniversityofScienceandTechnologyofChinain2008.AfterthathejoinedtheUniversityofFloridaasaPh.D.studentfortheDepartmentofComputerInformationandScienceEngineering.HisadvisorisDr.ShigangChen,andhisresearchinterestsincludeRFIDtechnologiesandinternettrafcmeasurement. 122



PAGE 1

nrr r rr rr r !"#$%& '(!(")(#*("+ & ,-.#/r 01/20++ & 3(/444/#2#)+ r& +r& + & 5678559+ r& :+& ;%$%++