Samir Goel

1

• Samir Goel Tomasz Imielinski
• PI Badri Nath
• Site Visit Briefing
• August 2001
• http//www.cs.rutgers.edu/dataman/webdust
• badri_at_cs.rutgers.edu
• Co-PIs Tomasz Imielinski, Rich Martin

2
Prediction-based Monitoring in Sensor Networks

• Many useful applications in sensor networks
  require monitoring sensors
• Example intrusion detection, target tracking,
  etc.
• We examine highly energy-efficient mechanisms for
  monitoring in sensor networks

3
Motivation

• Sensors run on batteries with limited lifetime
• Battery lifetime is improving at a slow rate
• 2-3 per year

A sensor lives to say only so much

• Energy-efficient Mode of operation seems to be
  the best alternative
• Transmissions drain energy
• Batteries last anywhere from 30 hrs to a year
  TinyOS
• Ideally, we want to perform monitoring with
  minimum number of transmissions from sensors

4
Approach

• Key Observation
• Sensors in close proximity are likely to have
  correlated readings
• Example Spatio-temporal Correlation

S1
S2
S3
S4
5
Approach

• Key Observation
• Sensors in close proximity are likely to have
  correlated readings
• Example Spatio-temporal Correlation

S1
S2
S3
S4
6
Approach

• Key Observation
• Sensors in close proximity are likely to have
  correlated readings
• Example Spatio-temporal Correlation

S1
S2
S3
S4
7
Approach

• Key Observation
• Sensors in close proximity are likely to have
  correlated readings
• Example Spatio-temporal Correlation

S1
S2
S3
S4
8
Approach

• Key Observation
• Sensors in close proximity are likely to have
  correlated readings
• Example Spatio-temporal Correlation

S1
S2
S3
S4
9
Approach

• Key Observation
• Sensors in close proximity are likely to have
  correlated readings
• Example Spatio-temporal Correlation
• Use correlation to predict readings of sensors
• A sensor need not transmit an expected reading
• Correlation doesnt imply redundancy

S1
S2
S3
S4
10
New Paradigm of Operation PREMONition

• Mode of operation
• Base computes a prediction and sends it to the
  sensor
• Sensor transmits its reading only when it is
  different from the predicted one

Its not news if one can predict it
S1
S2
S3
S4
11
PREMON paradigm in action

• Mode of operation
• Base computes a prediction and sends it to the
  sensor
• Sensor transmits its reading only when it is
  different from the predicted one

Its not news if one can predict it
S1
S2
S3
S4
12
PREMON paradigm in action

• Mode of operation
• Base computes a prediction and sends it to the
  sensor
• Sensor transmits its reading only when it is
  different from the predicted one

Its not news if one can predict it
S1
S2
S3
S4
13
PREMON paradigm in action

• Mode of operation
• Base computes a prediction and sends it to the
  sensor
• Sensor transmits its reading only when it is
  different from the predicted one

Its not news if one can predict it
S1
S2
S3
S4
14
PREMON paradigm in action

• Mode of operation
• Base computes a prediction and sends it to the
  sensor
• Sensor transmits its reading only if it differs
  from the predicted one

Its not news if one can predict it
S1
S2
S3
S4
15
PREMON paradigm in action

• Mode of operation
• Base computes a prediction and sends it to the
  sensor
• Sensor transmits its reading only when it is
  different from the predicted one

Its not news if one can predict it
S1
S2
S3
S4
16
PREMON Key characteristics

• Trades computation for communication
• Cost(computation) ltlt Cost(communication)
• Applicable whenever correlation (temporal,
  spatial, or spatio-temporal) exists
• We will focus on spatio-temporal correlation in
  this presentation

17
Computing Predictions Key Observations

• Snapshot of the state of sensor network may be
  visualized as an image
• Monitoring may be seen as watching a
  video of
  sensed values
• Techniques from MPEG-2 video coding
  
  standard may be used in PREMON

18
MPEG-2 Motion-compensation

• Block-matching algorithm
• Resilient to errors
• Applicable to any type of image

Frame1
Frame2
19
PREMON Motion-prediction

• Interpolate the readings of sensors at regular
  grid points
• Apply block-matching algorithm to compute
  motion-vectors
• Translate motion-vectors into motion-predictions

Frame1
Frame2
Frame3
20
Initial Results

• Experimental Setup

• One-dimensional version of the problem
• Base-station code fully resides in a mote
• Experimental Results

Summary of Results - Case3 performs 5 times
better than case1 - Case3 performs 28 better
than case2
21
Analysis

• Some constants
• Cost of transmission/bit 1 µJ, Cost of
  reception/bit 0.5 µJ
• Cost of computing 0.8 µJ per 100 instructions
• Update size 11 bytes (Tu 88 µJ, Ru 44 µJ)
• Prediction size 9 bytes (Tp 72 µJ, Rp 36
  µJ)
• BS makes a constant-value prediction after 2
  frames
• Every object movement causes 3 transmissions
• Let, s be the success ratio for predictions
• Cc be the cost of computing
  predictions

22
Analysis contd

• Cost in Premon mode (Case3)
• s x energy consumed when the prediction succeeds
  (1-s) x energy consumed when the prediction
  fails
• C3 s(2Cc 72 36) (1-s) (2Cc 72 36
  3(88 44))

• Cost in Case2 (transmit whenever the readings
  change)
• C2 3(88 44))
• When is C3 ? C2 ?
• s ? (2Cc 108)/396
• Size of code computing predictions is 1200
  instructions
• Cc (21200)energy/instruction 19.2 µJ
• s ? 0.394
• Cc ? (396s - 108)/2
• For 75 success rate
• Cc ? 94.5 µJ (enough for performing 11,800
  instructions)

23
Conclusions

• Prediction-based monitoring paradigm can
  significantly increase energy efficiency
• Monitoring of sensor data may be visualized as
  watching a video and MPEG-2 algorithms may be
  adapted for generating predictions

24
Status of Work / Plan

• Completed by July01
• Defined the conceptual framework for
  Prediction-based monitoring and worked out the
  mechanisms
• Implemented a small prototype with 3 sensors
  lined along a corridor
• Extended the prototype to work with 4 sensors
  lined along a corridor
• Not a trivial extension!
• Plan for next year (Aug01 - July02)
• Extend the prototype to work with two-dimensional
  motion
• Introduce learning in Premon

• Simulation study of performance of PREMON in
  large scale sensor networks

25
Plan

• Plan for Aug02 July03
• Change duty cycles of sensors based on predicted
  pattern of activity
• Extract pattern at larger time-scale and use
  these to drive the short-term prediction process

26
Information

• http//www.cs.rutgers.edu/dataman/webdust
• Papers
• Samir Goel and Tomasz Imielinski,
  Prediction-based Monitoring in Sensor Networks
  Taking Lessons from MPEG. Technical Report
  DCS-TR-438, Department of Computer Science,
  Rutgers University, June 2001. Submitted for
  publication in Computer Communication Review,
  Special Issue on Wireless Extensions to the
  Internet, October 2001.
• Samir Goel and Tomasz Imielinski,
  Prediction-based Monitoring in Sensor Networks
  Taking Lessons from MPEG. DIMACS Workshop on
  Pervasive Networking, May 21, 2001.
• Tomasz Imielinski and Samir Goel, DataSpace -
  querying and monitoring deeply networked
  collections in physical space, IEEE Personal
  Communications Magazine, Special Issue on
  "Networking the Physical World", October 2000.
• E-mail gsamir_at_CS.Rutgers.edu, tomasz_at_CS.Rutgers.e
  du