Upload
yackeline-ternera-pertuz
View
237
Download
0
Embed Size (px)
Citation preview
8/15/2019 AdaptiveRFID Hamid
1/32
Adaptive Cleaning forAdaptive Cleaning for
RFID Data StreamsRFID Data Streams
Shawn Jeffery Minos Garofalakis Michael Franklin
C !erkeley Intel Research !erkeley C !erkeley
"resented #y$ %amid %aidarian Shahri
8/15/2019 AdaptiveRFID Hamid
2/32
Where Are We? Look at the Signs!Where Are We? Look at the Signs!
8/15/2019 AdaptiveRFID Hamid
3/32
Looking at Signs – Before Jumping InLooking at Signs – Before Jumping In
& S. Chaudhuri, U. aa", #An $%er%ie& ofata Warehousing and $LA' (e)hno"og,#SIGMOD Record , *++. -/ )itations
& W and information integration
& 0ata )"eaning1 term pu2"i)i3ed Identified its importan)e in integration
& 45tensi%e resear)h fo""o&ed
8/15/2019 AdaptiveRFID Hamid
4/32
6LB 7*6LB 7*
& Session 8*79 A(A :UALI(; < CL4A=I=>
& e)"arati%e data )"eaning9 "anguage, mode", anda"gorithms
Helena Galhardas (INRIA Rocquencourt), Daniela Florescu(Propel), Dennis Shasha (N!), "ric Si#on, and $ristian% Au&ustin Saita (INRIA Rocquencourt)
& 'otters &hee"9 an intera)ti%e data )"eaning sstem'iashan*ar Ra#an and +oseph M Hellerstein(!ni-ersit o. $ali.ornia at /er*ele)
& Update propagation strategies for impro%ing the @ua"it ofdata on the We2
Ale0andros 1a2rinidis and Nic* Roussopoulos (!ni-ersito. Marland)
8/15/2019 AdaptiveRFID Hamid
5/32
ata C"eaning 're%ious Work 7ata C"eaning 're%ious Work 7
& amid aidarian Shahri, S.. Shahri, 04"iminating up"i)ates inInformation Integration9 An Adapti%e, 45tensi2"e Drame&ork,# I"""Intelli&ent Sste#s, 6o". 7*, =o. E, 7.
8/15/2019 AdaptiveRFID Hamid
6/32
'utting (hings into Conte5t'utting (hings into Conte5t
& ata )"eaning re@uired after integration
=o unified standard a)ross sour)es
=$W9 sensorFhard&are errors ine%ita2"eG
resear)h opportunit
& ata mode"ing HAmo" eshpande
An i#portant use )ase is )"eaning
8/15/2019 AdaptiveRFID Hamid
7/32
6LB 7 – (hree &eeks ago6LB 7 – (hree &eeks ago
& 8esear)h Session E9 Sensor ata Hdedi)ated to )"eaning!
& (it"e9 Adapti%e C"eaning for 8DI ata Streams
Authors9 Sha&n 8. Jeffer, Kinos >arofa"akis, Ki)hae" J.Drank"in
& (it"e9 A eferred C"eansing Kethod for 8DI ata Ana"ti)s
Authors9 Jun 8ao, Sangeeta orais&am, eta" (hakkar, Latha S.Co"2
& (it"e9 $n"ine $ut"ier ete)tion in Sensor ata Using =on'arametri) Kode"s
Authors9 Sharmi"a Su2ramaniam, (hemis 'a"pana, imitris'apadopou"os, 6ana a"ogeraki, imitrios >unopu"os
8/15/2019 AdaptiveRFID Hamid
8/32
8DI9 8adio Dre@uen) Ientifi)ation8DI9 8adio Dre@uen) Ientifi)ation
8/15/2019 AdaptiveRFID Hamid
9/32
8DI data is dirt8DI data is dirt
A simple experiment:
•2 RFID-enabledshelves
•10 static tags•5 mobile tags
8/15/2019 AdaptiveRFID Hamid
10/32
8DI ata C"eaning8DI ata C"eaning
Time
Rawreadings
Smoothedoutput
& 8DI data has man dropped readings& (pi)a"", use a s#oothin& .ilter to interpo"ate
S!"# distinct tag$id
FR%& RFID$stream 'RA() *5 sec+,)R%. / tag$id
S!"# distinct tag$idFR%& RFID$stream 'RA() *5 sec+,)R%. / tag$id
/t how to set the si3e
o4 the indo6
/t how to set the si3e
o4 the indo6
Smoothing Filter
8/15/2019 AdaptiveRFID Hamid
11/32
Windo& Si3e for 8DI SmoothingWindo& Si3e for 8DI Smoothing
Fido moving Fido resting
Small windowSmall window
RealityRealityRaw readingsRaw readings
Large windowLarge window
Need to balance completeness vs.capturing tag movement
Need to balance completeness vs.capturing tag movement
8/15/2019 AdaptiveRFID Hamid
12/32
(ru" e)"arati%e Smoothing(ru" e)"arati%e Smoothing
& 'ro2"em9 &indo& si3e nonde)"arati%e App"i)ation &ants a )"ean stream of data
Windo& si3e is ho3 to get it
& So"ution9 adapt the &indo& si3e inresponse to data
8/15/2019 AdaptiveRFID Hamid
13/32
ItinerarItinerar
& Introdu)tion9 8DI data )"eaning
& A statisti)a" samp"ing perspe)ti%e
& SKU8D 'ertag )"eaning
Ku"titag )"eaning
& $ngoing &ork& Con)"usions
8/15/2019 AdaptiveRFID Hamid
14/32
A Statisti)a" Samp"ing 'erspe)ti%eA Statisti)a" Samp"ing 'erspe)ti%e
& 4e Insi&ht98DI data ≈
random samp"e of present tags
& Kap 8DI smoothing to a samp"inge5periment
8/15/2019 AdaptiveRFID Hamid
15/32
8DIMs >or etai"s8DIMs >or etai"s
Epoch TagID ReadRate
0 1 .9
0 2 .6
0 3 .3
#ag 1
#ag 2
#ag 7
#ag 8
Antenna & readerTags
! " # $ % ' ( )*
Read "9cleEpoch;
+For Alien readers,
Tag List
8/15/2019 AdaptiveRFID Hamid
16/32
8DI Smoothing to Samp"ing8DI Smoothing to Samp"ing
8DI Samp"ing
8ead ))"e Hepoch Samp"e tria"
8eading Sing"e samp"e
Smoothing &indo& 8epeated tria"s
8ead rate 'ro2a2i"it of in)"usion Hpi
(o se sampling theor9 to drive adaptation<
8/15/2019 AdaptiveRFID Hamid
17/32
SKU8DSKU8D
& Statistical S#oothing for ! nreliable RF ID Data& Adapts window #ased on statistical properties& Mechanisms for$
& "er'tag and m(lti'tag cleaning
8/15/2019 AdaptiveRFID Hamid
18/32
'er(ag Smoothing9 Kode" and'er(ag Smoothing9 Kode" and
Ba)kgroundBa)kground
& Use a 2inomia" samp"ing mode"
Time
+epochs,
pi
!
*
Smoothing -indow
wi ernoulli trials
piavg
Si
+Read rateo/ tag i,
! " # $ % ' ( )*
8/15/2019 AdaptiveRFID Hamid
19/32
'er(ag Smoothing9 Comp"eteness'er(ag Smoothing9 Comp"eteness
& If the tag is there, read it &ith high pro2a2i"it Want a "arge &indo&
pi
!
*
Reading with alow pi
0pand the window
Time+epochs,! " # $ % ' ( )*
8/15/2019 AdaptiveRFID Hamid
20/32
'er(ag Smoothing9 Comp"eteness'er(ag Smoothing9 Comp"eteness
0pected epochs
needed to read
-ith probability
!1
2esired windowsi3e /or tag i
= δ
1ln*
1
avg i
i
pw
8/15/2019 AdaptiveRFID Hamid
21/32
'er(ag Smoothing9 (ransitions'er(ag Smoothing9 (ransitions
& ete)t transitions as statisti)a""signifi)ant )hanges in the data
pi
!
*
Statistically signi/icant di//erence
Flag a transition andshrin4 the window
The tag has li4ely le/t bythis point
Time+epochs,! " # $ % ' ( )*
8/15/2019 AdaptiveRFID Hamid
22/32
'er(ag Smoothing9 (ransitions'er(ag Smoothing9 (ransitions
5 e0pectedreadings
6s the di//erence 7statistically signi/icant89
5 observedreadings
)1(**2|*||| avg
i
avg
ii
avg
iii p pw pwS −>−
&Statisti)a"" signifi)antStatisti)a"" signifi)ant
8/15/2019 AdaptiveRFID Hamid
23/32
SKU8D in A)tionSKU8D in A)tion
Fido moving Fido resting
S:;RFS:;RF
0periments with real and simulated datashow similar results
8/15/2019 AdaptiveRFID Hamid
24/32
Ku"titag C"eaningKu"titag C"eaning
& Some app"i)ations on" need aggregates
4.g., )ount of items on ea)h she"f onMt need to tra)k ea)h tag!
& Use statisti)a" me)hanisms for 2oth5 Aggregate )omputation
Windo& adaptation
8/15/2019 AdaptiveRFID Hamid
25/32
Aggregate ComputationAggregate Computation
& –estimators Hor%it3(hompson
& Count9
& 'Ntag i seen in a &indo& of si3e &O9
Use sma"" &indo&s to )apture mo%ement
Use the estimator to )ompensate for "ost readings
∑ ∈∧
=wS iw N
ι
π
1
wavg
ii p )1(1 −−=π
8/15/2019 AdaptiveRFID Hamid
26/32
Windo& AdaptationWindo& Adaptation
& Upper 2ound &indo& simi"ar to pertag
& 0(ransition1 2ased on %arian)e &ithin
su2&indo&s
=
δ
1ln*
1avg
p
w
"ont
(
(+
Time+epochs,! " # $ % ' ( )*
+
>
∧∧
'Var Var 2ww
N N
8/15/2019 AdaptiveRFID Hamid
27/32
Ku"titag S)enarioKu"titag S)enario
8/15/2019 AdaptiveRFID Hamid
28/32
$ngoing Work9 Spatia" Smoothing$ngoing Work9 Spatia" Smoothing
& With mu"tip"e readers, more )omp"i)ated
Reinforcement
A9 9 A ; 9 A 9
Arbitration
A9 two readers per room
8/15/2019 AdaptiveRFID Hamid
29/32
Beond 8DIBeond 8DI
& estimator for other aggregates
Use SKU8D for sensor net&orks
& Use SKU8D in genera" streaming sstemsHe.g., (e"egraphC:
8emo%e RANGE )"ause from C:L
%ther sensor data%ther sensor data
%ther streaming data%ther streaming data
8/15/2019 AdaptiveRFID Hamid
30/32
8e"ated Work8e"ated Work
& Commer)ia" 8DI midd"e&are Smoothing fi"ters9 need to set smoothing &indo&
& 8DIre"ated &ork
8ao et a"., StreamC"ean9 )omp"ementar Inte" Seatt"e, iDi, 4S'9 stati) &indo& si3e
& BB:, Kau%eB ea%&eight, mode"2ased
SKU8D is nonparametri), samp"ing2ased
& Statisti)a" fi"ters Hdigita" signa" pro)essing < B =on"inear digita" fi"ters inspired SKU8D design
8/15/2019 AdaptiveRFID Hamid
31/32
Con)"usionsCon)"usions
•
8/15/2019 AdaptiveRFID Hamid
32/32
(hanks!(hanks!
:uestions?