Adaptive Sampling for Sensor Networks

1
Adaptive Sampling for Sensor Networks

• Ankur Jain?and Edward Y. Chang
• University of California, Santa Barbara
• DMSN 2004

2
Outline

• Sampling in sensor networks
• Adaptive sampling using Kalman Filter
• Problem formulation
• Results

3
Sampling in Sensors

• Sampling Interval (SI) time interval between
  successive measurements
• Sensitive to streaming data characteristics,
  query precision and available resources
• Over-sampling comes at increased resource usage
• CPU at the sensor and the central server
• Network Bandwidth within the sensor network
• Power Usage at the sensor

4
Examples

• Habitat Monitoring Animal activity
• Higher bandwidth to sensors reporting
  interesting events
• Unusual changes in temperature, sound levels
• Video Surveillance Parking Lot
• Higher rate video capturing in area experiencing
  unexpected traffic pattern
• Random swirling, speeding

5
Related Work

• Network Contention
• Considers network contention before putting data
  on the network channel
• Better delivery rate at the server
• Stochastic Estimation
• Adapts to input data characteristics using
  stochastic models
• Does not consider multiple sensors scenario

6
Modeling Streaming Data Characteristics

• A Kalman Filter (KF) is used by each sensor to
  estimate expected values (value at the next
  measurement)
• Estimation error (ER) from KF is used to
  quantify streaming data characteristics
• High error compensated by lower SI

7
The KF cycle
Measurement from the sensor
Measurement Update (Correct)
Time Update (Predict)
Estimation Error (ER)
8
Adaptive Sampling

• All sensors stream updates to a central server
• ER is calculated at each measurement
• Based on ER, the sensors can adjust the sampling
  interval within a specified range SIR (Sampling
  Interval Range)
• Beyond the range the sensor requests the server
  for lower sampling interval (more bandwidth)
• The server allocates bandwidth based on available
  resources

9
Sensor Side

• No server mediation required as long as the
  desired change in Sampling Interval (SI) is
  within SIR
• SI last last SI received from the server
• SI desired desired SI to reduce ER
• High activity streams can be captured at low SI
  avoiding delays due to server response or network
  congestion

10
Sensor Side

• New SI is proportional to estimation error from
  the KF over a sliding window of sizeW
• SI new desired SI
• SI current current SI
• ? user parameter (max. change in SI)
• f fractional change in ER over sliding window
• If SI new is out of range, a new SI is
  requested from the server
• ?SI change in SI requested

11
Server Side

• The server puts requests in a queue with 5
  attributes
• Fractional Error (f) fractional error at the
  sensor
• Request (Req) change in SI requested
• History (h) age of the request in the queue
• Grant (g) amount by which the request has been
  satisfied
• Query Weight (w) Weight from the query
  processor
• The server forms an optimization problem such
  that A is the amount granted and Ravail is the
  available resource

12
Experiments

• Oporto simulator used to obtain trajectories of
  moving shoals
• One sensor per shoal (12 Shoals)
• 3000 tuples at each sensor
• Results compared with uniform sampling approach
• Effective Resource Utilization (ERU) ?

? mean fractional error between real and actual
trajectory m fraction of messages exchanges
between sensors and server
13
Results ERU vs. Number of Sources
14
Results ERU vs. Sliding Window Size
15
Future Work

• Extension to multi hop sensor networks
• Application of other estimation models (particle
  filters)
• Dynamic SIRs
• Development of better algorithms to reduce
  message overheads

16
Thank you !