Adaptive Data Aggregation for Wireless Sensor Networks
S. Jagannathan
Rutledge-Emerson Distinguished Professor
Department of Electrical and Computer Engineering
Professor of Computer Science
Missouri University of Science and Technology
Rolla, MO 65409.
1Research performed by Priya Kasirajan is thankfully acknowledged
Agenda• Introduction• Background• Challenges• Proposed Methodology• Results and Discussion• Hardware results
• Conclusions and Future work
2
• Why compression?– Reduction in amount of data transmitted– Reduction in energy consumption– Improvement in network lifetime
• Compression vs Aggregation– Data condensed at the source
node– Aggregation implies data from spatially separated sensors combined
statistically using min, avg, max, count, sum– Need location or node ID
Node
Clusterhead
Introduction
3
Background
4
• Survey of data aggregation (Rajagopalan and Varshey, 2006)– Chain based data aggregation, tree based, PEDAP, Grid based, Network flow based, network
correlated data aggregation, QoS-aware aggregation
• Quantization
– Lossy compression scheme
– Quantization error is proportional to step size
– Step size is dependent on dynamic range
• Adaptive Differential Pulse Code Modulation (ADPCM)
– Quantize difference between actual sample and estimated sample
– Exploits the correlation between adjacent samples to reduce bit rate and to achieve compression.
• Real world sensor data with multiple modalities does not always boast correlation and linear relationship
Challenges
• Data compression/aggregation can be a complex nonlinear process
• Nonlinear processing is computationally more intensive
• Data reconstruction can be involved– Location aware or context aware– Node ID
• Performance guarantees in terms of distortion, compression ratio, energy efficiency, hard to show
5
Proposed MethodologyChannel
e(k)
Some y(k)
y(k)
Quantizer
EstimatorEstimator EncoderEncoder DecoderDecoder EstimatorEstimator
Analytical ResultsTheorem 1 (Estimator-Ideal Performance): In the ideal case with no
reconstruction errors and noise present, the estimation error approaches to zero asymptotically while the parameter estimation error vector is bounded.
Theorem 2 (Estimator Performance—General Case): Let the hypothesis presented in Theorem 1 hold and if the functional reconstruction error is bounded, then estimation error is bounded while the parameter errors are also bounded.
Analytical Results (contd.)
Theorem 3 (NADPCMC Distortion): If the estimator reconstruction and quantization errors are considered bounded, then the distortion at the destination is bounded. On the other hand in the absence of estimator reconstruction and quantization errors, the distortion is zero.
Theorem 4 (NADPCMC Performance): The compression ratio, defined as the ratio of the amount of uncompressed data to the amount of compressed data, is greater than one. Moreover, the proposed scheme will render energy savings.
8
Simulation Results
• River Discharge Data• Audio Data• Geophysical Data
FLoating point Operations Per Second – FLOPSNADPCMC encoding
7050 FLOPS 1.224 micro joules
NADPCMC decoding7425 FLOPS 1.289 micro joules
XBee radio – transmit power – 1 mW for 30 m
Energy Consumption
River Discharge Data
0 20 40 60 80 100 120 140 160 1800
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Iteration
Am
plit
ud
e
NADPCMC with 8 bit encoded error
Original
Decoded
0 20 40 60 80 100 120 140 160 180-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Iteration
Re
con
stru
ctio
n e
rro
r
Total error in NADPCMC reconstruction
8 bit error
6 bit error
Time Time
River Discharge Data (contd.)
11
MethodCompression
ratio
Energy savings at
nodes
Energy savings at CH
Distortion Overhead
Huffman 1.453 NA 31.177% NA 480 bytes
Differential Huffman 1.642 21.56% 39.099% NA 480 bytes
Scaling and
approximation1.137 13.65% 11.65% 0.0657% 0
Scaling and 9 bit
quantization1.778 43.76% 43.76% 0.943% 0
Scaling and 8 bit
quantization2.000 50% 50% 2.0685% 0
Scaling and5 bit
quantization3.200 68.75% 68.75% 16.451% 0
Linear ADPCM 2 50% 50% 13.72% 0
NADPCMC with 8 bit
encoding1.9459 48.61% 48.61% 2.65% 10 bytes
NADPCMC with 6 bit
encoding2.5487 60.76% 60.76% 6.08% 10 bytes
Audio Data
0 0.5 1 1.5 2 2.5 3 3.5 4
x 104
-0.025
-0.02
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
0.02
0.025
Iteration
To
tal r
eco
nst
ruct
ion
err
or
NADPCMC with 8 bit encoded error
0 0.5 1 1.5 2 2.5 3 3.5 4
x 104
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
IterationT
ota
l re
con
stru
ctio
n e
rro
r
NADPCMC with 6 bit encoded error
Time Time
Audio Data (contd.)
13
Method Data rateCompression ratio
Energy savings at nodes
Distortion Overhead
Scaling and 8 bit
quantization
64kbps 2 50% 10.59% NA
Scaling and 6 bit
quantization
48kbps 2.67 62.5% 46.28% NA
5 bit linear ADPCM 40kbps 3.199 68.74% 11.37% NA
4 bit linear ADPCM 32kbps 4 75% 23.14% NA
3 bit linear ADPCM 24kbps 5.332 81.25% 28.45% NA
2 bit linear ADPCM 16kbps 8 87.5% 35.86% NA
NADPCMC with 8 bit
encoding
64kbps 1.9992 49.98% 2.04% 20 bytes
NADPCMC with 6 bit
encoding
48kbps 2.6653 62.48% 6.16% 20 bytes
NADPCMC with 4 bit
encoding
32kbps 3.997 74.98% 14.44% 20 bytes
0 100 200 300 400 500 600 700 800 900 1000-0.03
-0.025
-0.02
-0.015
-0.01
-0.005
0
Iteration
To
tal r
eco
nst
ruct
ion
err
or
NADPCMC with 8 bit encoded error
Geophysical Data Performance
14
0 100 200 300 400 500 600 700 800 900 1000-0.14
-0.12
-0.1
-0.08
-0.06
-0.04
-0.02
0
Iteration
To
tal r
eco
nst
ruct
ion
err
or
NADPCMC with 6 bit encoded error
Time Time
MethodCompression ratio
Energy savings at nodes
Distortion Overhead
Scaling and 8 bit quantization 2 50% 4.36% 0
Scaling and 6 bit quantization 2.667 62.5% 13.42% 0
Linear ADPCM 2 50% 35.87% 0
NADPCMC with 8 bit encoding 2 50% 1.02% 20 bytes
NADPCMC with 6 bit encoding 2.667 62.5% 4.22% 20 bytes
Aggregation using NADPCMC
• 8 bit NADPCMC at all source nodes• 6 bit NADPCMC at CH 1, 2 and 3 –
61.34% savings – 1.90%• 4 bit NADPCMC at CH 1, 2 and 3 –
73.61% savings - 6.10%• 4 bit NADPCMC at CH 5 – 74.54%
savings –– Synthetic data: 7.01%– River discharge data: 4.83%– Audio data: 6.09%
15
Hardware Implementation
16
• Compression ratio – 1.846• Energy savings – 45.83%• Distortion – 1.67%
• Compression ratio – 2.526• Energy savings – 60.42%• Distortion – 4.60%
Nano Sensor Data Performance
17
0 100 200 300 400 500 600 700 8000
0.5
1
1.5
2
2.5
3
3.5
Time
Am
plit
ud
e
Compression and Aggregation
Original
6 bit Aggregation4 bit Aggregation
0 20 40 60 80 100 120 140 160 1800
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Time
Am
plit
ud
e
Compression and aggregation
Original
6 bit aggregation4 bit aggregation
Data
generation
rate
Transmissio
n rate
Compress
ion ratio
Energy
savings
Distortion
Uncompress
ed data
2.56 kbps 3.424 kbps NA NA NA
Compressed
data
2.56 kbps 1.712 kbps 1.8751 46.67% 0.78% for sensor data
0.81% for river discharge data
Compressed
and
aggregated
data – 6 bit
NADPCMC
2.56 kbps 1.284 kbps 2.3250 56.99% 3.58% for sensor data
2.78% for river discharge data
Compressed
and
aggregated
data – 4 bit
NADPCMC
2.56 kbps 856 bps 3.5002 71.43% 8.21% for sensor data
10.90% for river discharge data
Conclusions
• Data aggregation process is nonlinear and must be location/self-aware for enhanced performance
• NADPCMC addresses nonlinear issues in data and performs well for different sensor modalities.
• Aggregation is achieved through iterative compression.
• Performance depends on number of aggregation levels and Quantizer resolution.
• Network size does not impact performance.
• Future work involves evaluation of the proposed scheme for larger size networks with different types of data by considering latency, life time and security