Removing Redundancy from Wireless Sensor Networks

1
Removing Redundancy from Wireless Sensor Networks

• Dragan Petrovic

Jim Chou Kannan Ramchandran
Rahul Shah Jan Rabaey
2
Overview

• Setup - Motivation
• Distributed Source Coding
• Theory
• Practical Codes
• Application to Sensor Networks
• Routing with Aggregation
• Results
• Conclusion

3
Scenario

• Network consists of many sensors, few
  controllers, and few actuators
• Sensors give their measurements to controllers,
  which process them and make decisions

4
Data Reports

• Payload has redundancy
• because nearby sensors
• have correlated readings
• Headers are overhead
• of communication setup
• control

5
Overview

• Setup - Motivation
• Distributed Source Coding
• Theory
• Practical Codes
• Application to Sensor Networks
• Routing with Aggregation
• Results
• Conclusion

6
Problem Base Case
Y

• Suppose X, Y correlated
• Y available at decoder but not at encoder of X
• How to compress X close to H(XY)?

X
7
Encoding with fidelity criterion
Rate distortion with side information (Wyner-Ziv,
78)

• X and Y are correlated,
• continuous alphabet sources.
• Y is available only to decoder.

X
Decoder
Encoder
Y
X aY N N correlation noise If N is
Gaussian, can get same performance by knowing
only statistics of N at encoder as when Y is
known at encoder. Performance is measured as MSE
for a given transmission rate R.
8
Basic Concept
Consider a simple 8-level scalar quantizer
Partition
r r r
r
0 2 4
6
r r r
r
2 cosets
1 3
5 7
Encoder Send index of coset containing quantized
outcome Decoder Decode side information Y in
given coset
9
r r r r r r
r r
0 1 2 3 4
5 6 7
X
Y
10
General block diagram of DISCUS (Pradhan
Ramchandran 99)
Decoder
Encoder
Find quantization index using source
codebook
Source X
Find Coset of quantized codeword
I
I
Find codeword closest to S in encoded
coset
Optimally estimate source
coset
Correlated source Y

• Interplay between source coding, channel coding
  and estimation theory can leverage latest
  advances in all areas (generalized coset codes
  Forney)
• 7-15 dB gains in reconstruction SNR over
  theoretically optimal strategies that ignore
  correlation for typical correlated sources.
• 3-5 dB gap from Wyner-Ziv bound (Gaussian case)
  using sample by sample quantization and encoding
  better performance with TCQ/TCM.

11
Challenges of Real World

• Theory says what is possible given the
  statistics.
• Codes exist which achieve bounds when statistics
  are known.
• How does one find the statistics?

Correlation Tracking Algorithms What codes to
use?
Practical Codes How is it Possible?
Theory What is Possible?
Real Systems
12
Setup

• Controller receives uncoded data from sensors
• Breaks them up into clusters s.t. nodes within
  cluster are highly correlated
• Tells each cluster what code-book to use

13
Correlation Tracking

• Sensor nodes measure X, controller node has Y

• Data gathering node needs to estimate number of
  bits, i, it needs from sensor nodes for X.

X aY N N correlation noise
14
Tree-Structured Code

• Depth of tree specifies number of bits used for
  encoding

D

• Path in the tree specifies the encoded value.
• Can tolerate 2i-1D of correlation noise using an
  ith level codebook

15
Simulations (correlation tracking)

• 15 sensors measuring temperature, light and
  humidity are arranged in star topology
• 3 clusters formed

Correlation Noise
Time
16
Simulations (energy savings)

• Each node has 0 energy savings 20 time (uncoded)
• Ave Temp Savings 40.8
• Ave Humidity Savings 17.6
• Ave Light Savings 6.75

Energy Savings ()
Time
17
Room for Improvement

• Encode more than one sample at a time
• Use a tighter bound (Chebyshev is too
  conservative)
• Larger clusters
• Non-stationary signals (e.g. acoustic)

18
Overview

• Setup - Motivation
• Source Coding with Side Information
• Theory
• Practical Codes
• Application to Sensor Networks
• Routing with Aggregation
• Results
• Conclusion

19
Motivation for Removing Header Redundancy

• In a sensor network, data rates are low packet
  header overhead is high
• Many sensors report back to controller
• Sending the information of many sensors in one
  packet reduces header overhead

20
Setup

• Network consists of many sensors, few
  controllers, and few actuators
• Sensors give their measurements to controllers,
  which process them and make decisions

21
Data Reports

• If sensors send their
• readings independently
• many headers sent
• unnecessarily

22
Routing with Aggregation

• Would like to do data aggregation along the way

Controller Sensors
23
Setup

• Network consists of many sensors, few
  controllers, and few actuators
• Sensors give their measurements to controllers,
  which process them and make decisions

24
Regions

• Controller defines a cuboid in space
• Regions can be chosen based on correlation within
  region
• Sensors within that region send their
  measurements to the controller every few minutes

Controller Sensors
25
Interest Packets

• Controller sends an interest packet defining
  the region and specifying frequency of sensor
  reports
• Nodes outside the region forward the packet and
  update path-cost within packet

Controller Sensors
26
Interest Packets

• Controller sends an interest packet defining
  the region and specifying frequency of sensor
  reports
• Nodes outside the region forward the packet and
  update path-cost within packet

Controller Sensors
27
Interest Packets

• Controller sends an interest packet defining
  the region and specifying frequency of sensor
  reports
• Nodes outside the region forward the packet and
  update path-cost within packet

Controller Sensors
28
Interest Packets

• Controller sends an interest packet defining
  the region and specifying frequency of sensor
  reports
• Nodes outside the region forward the packet and
  update path-cost within packet

Controller Sensors
29
Interest Packets

• Controller sends an interest packet defining
  the region and specifying frequency of sensor
  reports
• Nodes outside the region forward the packet and
  update path-cost within packet

Controller Sensors
30
Interest Packets

• Controller sends an interest packet defining
  the region and specifying frequency of sensor
  reports
• Nodes outside the region forward the packet and
  update path-cost within packet

Controller Sensors
31
Border Nodes

• Each border node computes its cost for reaching
  controller
• The cost is a function of the energy required to
  reach the controller and energy left in nodes
  along the path
• There are several paths, so each border nodes
  cost is a weighted average of the cost of each
  path

Controller Sensors
Border Nodes
32
Flood Packets

• Each border node floods the region with a packet
  identifying itself as a border node and
  specifying its cost to reach the controller,
  along with original interest by the controller
• Other nodes in region use these packets to
  determine paths back to the border nodes

Controller Sensors
Border Nodes
33
Flood Packets

• Each border node floods the region with a packet
  identifying itself as a border node and
  specifying its cost to reach the controller,
  along with original interest by the controller
• Other nodes in region use these packets to
  determine paths back to the border nodes

Controller Sensors
Border Nodes
34
Flood Packets

• Each border node floods the region with a packet
  identifying itself as a border node and
  specifying its cost to reach the controller,
  along with original interest by the controller
• Other nodes in region use these packets to
  determine paths back to the border nodes

Controller Sensors
Border Nodes
35
Compute Schedule

• Each node in region uses the same deterministic
  function to compute a schedule of when to talk to
  each border node

Controller Sensors
Border Nodes
36
Reporting to Border Node

• During every round of reporting, the sensors send
  their information to the designated border node
  for that round
• Sensors far from the border node send their info
  before the sensors near the border node in order
  to allow packet aggregation at intermediate nodes

Controller Sensors
Border Node
37
Reporting to Border Node

• During every round of reporting, the sensors send
  their information to the designated border node
  for that round
• Sensors far from the border node send their info
  before the sensors near the border node in order
  to allow packet aggregation at intermediate nodes

Controller Sensors
Border Node
38
Reporting to Border Node

• During every round of reporting, the sensors send
  their information to the designated border node
  for that round
• Sensors far from the border node send their info
  before the sensors near the border node in order
  to allow packet aggregation at intermediate nodes

Controller Sensors
Border Node
39
Reporting to Controller

• The designated border node aggregates all the
  information sent to it by the sensors in the
  region
• It sends out a single packet, with a single
  header, containing all the information from the
  region

Controller Sensors
Border Node
40
Reporting to Controller

• The designated border node aggregates all the
  information sent to it by the sensors in the
  region
• It sends out a single packet, with a single
  header, containing all the information from the
  region

Controller Sensors
Border Node
41
Reporting to Controller

• The designated border node aggregates all the
  information sent to it by the sensors in the
  region
• It sends out a single packet, with a single
  header, containing all the information from the
  region

Controller Sensors
Border Node
42
Reporting to Controller

• The designated border node aggregates all the
  information sent to it by the sensors in the
  region
• It sends out a single packet, with a single
  header, containing all the information from the
  region

Controller Sensors
Border Node
43
Reporting to Controller

• The designated border node aggregates all the
  information sent to it by the sensors in the
  region
• It sends out a single packet, with a single
  header, containing all the information from the
  region

Controller Sensors
Border Node
44
Simulation

• Implemented Funneling algorithm in OpNet
  network simulator
• Controller queried a region with 15 sensors
• Sensors reported every 10 seconds with data
  aggregation along the way
• Measured the average number of sensor readings
  per packet transmittion

45
Topology
46
Results

• On average, 7 sensor readings per packet (i.e.
  per header)
• 85 reduction in the amount of energy spent on
  transmitting packet headers

47
Combining the Gains
a
(1-a)
Overall Savings a(DISCUS savings)
(1-a)(Funneling Savings)
48
Conclusion

• Sensor networks have a lot of built-in redundancy
  that saps the limited energy available.
• Distributed source coding can be used to reduce
  the redundancy of the sensor readings
• Routing with aggregation can be used to reduce
  the redundancy of communication set-up.