Cost Conscious Cleaning of Massive RFID Data Sets

1
Cost Conscious Cleaning of Massive RFID Data Sets

• Hector Gonzalez, Jiawei Han, Xuehua Shen
• University of Illinois at Urbana Champaign
• Department of Computer Science
• The Database and Information System Laboratory
• ICDE 2007

2
Outline

• Motivation
• RFID Data
• Error Sources
• Cleaning Methods
• Smoothing methods
• Rule based Methods
• DBN based methods
• Cost conscious cleaning
• Architecture
• Cleaning Methods
• Cleaning Sequence
• Cleaning Plan Induction
• Performance Study

3
Motivation

• The reliability of current RFID systems is far
  from optimal.
• Under a wide variety of environments more than
  50 of all tag readings are missed.
• The volume of data generated by RFID systems is
  huge
• A large retailer with hundreds of readers can
  generate thousands of tag readings every second.
• An accurate and efficient cleaning process is
  essential to the successful implementation of
  RFID technology.

4
Example

• A large retailer with RFID readers at warehouses,
  distribution centers, and store backrooms.
• A variety of factors impact correct tag
  detections
• Diverse reader/tag manufacturers, generation
• Moving (conveyor belts, doors) and static tags
  (shelfs).
• Different levels of RF noise caused by metal or
  water in the environment or in products.
• No single method can efficiently clean such a
  large volume of data, generated under such
  diverse circumstances.

5
RFID Data

• Readers conduct interrogation cycles at periodic
  invervals
• An RF signal is issued
• Tags awake and transmit via RF their EPC
  (electronic product code)
• A singulation protocol is used to prevent tag
  collisions
• In order to improve accuracy, during a read cycle
  multiple interrogation cycles are issued.
• Readings of the form (EPC, Reader Time) are
  generated at the end of each read cycle. Readings
  have extra information such as
• Total responses obtained during the read cycle
• Antenna used for detection, tag type, and signal
  strength

6
Error Sources

• There are two types of errors.
• False Negatives A tag that is present is not
  detected.
• False Positives A tag not present is detected.
• Why do we observe errors
• Collisions Multiple tags transmit
  simultaneously, or Multiple readers transmit
  simultaneously.
• Environment Interference Metal or water near the
  tag or reader cause RF interference.
• Physical Configuration The tag moves too
  quickly, or is located in a blind spot.
• Logical Errors A door reader detects tags that
  go nearby but not through.

7
Cleaning Methods

• A cleaning method M is a classifier that assigns
  a location to tag cases.
• A tag case is a tuple of the form
• (ltEPC, timegt,ltf1,f2,,fkgt)
• Where each f_i is a feature (e.g. tag type,
  signal strength)
• The label assigned to each tag case is of the
  form
• (ltEPC,timegt location, confidence)
• Where confidence is in the rage of 0.0 1.0 and
  indicates the level of certainty about the
  location.
• Terminal classifiers do not provide a confidence
  values

8
Fixed size smoothing window

• Fixed window smoothing
• The window is made up of the last k (fixed) read
  cycles.
• If there is any reading inside the window mark
  the tag as present
• Problems
• Difficult to define the best window size for
  different conditions
• Benefit
• Cheap method to apply, only requirement is to
  remember last k readings

Truth

Readings

Smooth

9
Adaptive window size smoothing

• Adaptive window size smoothing
• Change the size of the window according to the
  observed probability of tag detections
• Use a binomial model
• Let p, be the probability of detecting a present
  tag
• chose window size w, such that
• (1-p) w lt threshold
• Benefits We adapt the window size to the current
  conditions
• Problems
• It can be expensive to store and maintain p for
  every single tag in the system
• All readings inside the window have the same
  weight, better to put more weight on recent
  readings

10
Rule Based Cleaning

• Rules can be derived from the data or given by a
  domain expert.
• A door reader should only recognize tags with a
  bell shaped signal strength.
• The shelf reader should only recognize items that
  stay there for more than 5 minutes
• Rules can be derived from an RFID warehouse
• Flowgraphs can be used to complete missing
  readings and to decide location conflicts

11
DBN Based Cleaning

• We can model tag detections using a dynamic but
  hidden process
• Tag readings correspond to noisy observations on
  the true, but hidden, location of the tag
• We dynamically update our belief on tag location
  based on the sequence of observations
• More recent observations weight higher on our
  belief
• DBNs allow us to differentiate between the
  following two cases

No question, detect tag !!!
Should we detect this tag???
12
DBN Structure
Transition Model

• We learn the transition and observation models
  from data
• DBNs can represent complex structures e.g.,
  variable for RF noise, and tag speed interacting
  with detection

Present t
Present t-1
hidden
Detect t
Detect t-1
Observed
Observation Model
13
Belief state updating
Belief Update Equation
Observation Model
Belief at time t1 given all evidence up to t1
Update to belief state given transition model
Belief state is updated dynamically, we do not
need to remember a window of observations, we
only need to keep the latest belief state.
14
Cost Conscious Cleaning

• Given a collection of cleaning methods, a set of
  labeled tag cases, and a cost model for each
  cleaning method, design an efficient cleaning
  plan that defines the conditions under which each
  cleaning method or sequence of cleaning methods
  should be applied

15
System Architecture
RFID Stream
Labeled Data
Cleaning Plan Induction
Apply Plan
Cleaning Methods
Clean Data
Online
Offline
16
Cost Model

• In order to apply a cleaning method to a tag case
  we need to incur a cost
• Classification cost cost in terms of cpu and
  storage of labeling each tag case.
• Amortized per tuple training cost Cost to train
  the cleaning method, e.g., in a DBN we need to
  learn the transition and observation models.
• Error cost
• When we make an error in deciding the location of
  a tag, a cost is incurred.
• Error cost can be a scalar or a function of the
  distance of the correct location to the predicted
  one, or even the price of the item.

17
Cleaning Sequence

• Ordered application of cleaning methods for a set
  M to a set of tag cases D
• SD,M Ms1 ? Ms2 ? ? Msk
• Apply Ms1 to the entire data set D
• Apply Ms2 to the cases that Ms1 failed to
  classify
• Apply Msk to the cases that every other method
  failed to classify
• The cost of applying a cleaning sequence C(SD,M)
  is the cost of applying each method as described
  plus the error cost on tag cases misclassified by
  every method
• Optimal cleaning sequence SD,M is the cleaning
  sequence with minimal cost among all possible
  cleaning sequences given D, and M

18
Cleaning Sequence Approximation
Step 1
M2
M1
M3
Step 2
Step 3
C(M1)1, C(M2)1.5, C(M3)0.5, Error 5 Accuracy
Adjusted Cleaning Cost
SD,M M1 ? M3 ? M2
C(SD,M) 1 0.520.5 0.431.5 0.195
19
Cleaning Plan

• Input
• D Set of labeled tag cases lt(EPC,time),(features
  )gt
• M Set of available cleaning methods
  M1,M2,,Mk
• C Cost model C(M1),C(M2),,C(Mk),C(Error)
• Output
• A decision tree that splits D according to
  feature values.
• For each leaf in the tree a cleaning sequence is
  defined.
• Application
• For each test case,
• use feature values to get to appropriate leaf.
• apply cleaning sequence defined in the leaf.

20
Available Features

• Tag features
• Communications protocol, Manufacturer, price,
  quality
• Detection history
• Reader features
• Number of antennas
• Protocol, price, vendor
• Location features
• type of area being covered (e.g. door, shelf,
  conveyor belt)
• Interference level (e.g. presence of metal or
  water)
• Item features
• Type of item, contents, price

21
Cleaning Plan Induction

• Use a traditional Top Down Induction of Decision
  Trees (TDIDT) algorithm
• Node splitting criteria
• Split nodes based on expected cost reduction

Cleaning sequence cost before the split
Average cost for each cleaning sequence after the
split.
22
Example
Labeled tag cases
Cleaning Plan
reader
door
shelf
Method pat Accuracy 100
yes
metal
no
Method fix_1 Accuracy 75
Method dbn Accuracy 100

• We use 3 cleaning methods fix_1, DBN, and pat
• The label is 1 if the tag is indeed present, and
  0 otherwise
• Each method predicts 1 for present, 0 for absent
• The cleaning plan selects when to apply each
  method, e.g. we should use fix_1 to clean cases
  from shelf readers when there is metal

23
Example of node splitting
C(fix_1)1.0,C(DBN)2.0, C(PAT)2.0,
C(Error)5.0 Cleaning sequence for D DBN ? pat
2.0 0.332.0 0.115.0 3.21
DNB ? fix 2.0 0.161.0 2.16
pat 2.0
Cost Reduction 3.21 (2.160.67 2.00.33)
1.11
24
Experimental Setup

• Data generator simulates a complex RFID system
  with multiple locations, readers, and tag types.
• The simulation is controlled by several
  parameters
• Item flow characteristics, paths traversed,
  predictable vs random movements
• Item speed
• Reader, tag, and item characteristics (e.g.
  protocol, manufacturer, item contents).
• Location characteristics and RF noise levels
• The simulation is run for a number of read
  cycles, at each cycle to probability of tag
  detection is a function of reader, item, tag, and
  location characteristics.

25
Cleaning methods

• We compare the performance of several cleaning
  methods
• DBN a dbn based cleaner
• var adaptive window smoothing
• fix_k fixed (size k) window smoothing
• Rule based methods
• pat a pattern recognition method based on signal
  strength shape
• maj used to resolve multi reader detection
  conflicts by majority voting
• Cost Model
• C(DBN)2.0, C(var)2.0, C(fix_1) 1.0,
  C(fix_3)1.4, C(pat)2.5, C(maj)1.4, C(Error)10
• All methods are terminal except for DBN which
  uses P(tag present) as confidence

26
Complex Setup I

• readers in low noise
• readers detecting far away tags
• readers with variable detection rates
• reader at a conveyor belt
• readers detecting tags with water and metal
• In some areas there can be conflict of multiple
  patterns detecting the same tag

27
Complex Setup II
28
Reader Setup
29
Noise Level
30
Conclusions

• A cost conscious cleaning framework for RFID data
  can increase accuracy at lower cost than any
  single cleaning method.
• The cleaning plan can be efficiently learnt from
  data by applying the idea of cleaning cost
  reduction to node splitting.
• DBN based cleaning methods capture the intrinsic
  dynamic behavior of tag detection, and deliver
  high accuracy at lower costs than smoothing
  window based techniques.

31
Thanks