LEAP: Localized Energy-Aware Prediction for Data Collection in Wireless Sensor Network

foamyflumpMobile - Wireless

Nov 21, 2013 (3 years and 8 months ago)

52 views

LEAP: Localized Energy
-
Aware
Prediction for Data Collection in
Wireless Sensor Network

Hongbo Jiang

Shudong Jin


Mobile Ad Hoc and Sensor Systems(MASS), 2008

Outline


Introduction


Model selection


Prediction model


Energy
-
efficient algorithm


Algorithm design


Experiment results


Conclusion

Introduction


Nowadays, sensors’ processing power and compacity have been
increased. Sophisticated algorithm on sensors like
predictor

become possible.



A predictor use past input values from sensors to perform
prediction, implys sensors do not need to transmit the data if
they differ from a predicted value by less than a threshold



Simple prediction way is that all sensor nodes send data to base
station and only
base station

do the predictor training and
prediction, despite their increasing computing capacity. But it
cost high transmission energy, wireless bandwidth and potential
high latency.

Introduction

-

localized prediction


Clustering
-
based localized prediction divide the network into
clusters and use a cluster
-
head(also a sensor node) to maintain
each cluster member’s history data. Localized prediction may
highly energy
-
efficient due to the reduction of routing path.



There is a trade off between communication and computation
since prediction computation cost is not negligible in localized
prediction.

Base
station

Localized Energy
-
Aware Prediction
(LEAP)


Localized Energy
-
Aware Prediction (LEAP)


Cluster head


receive data that selective reported by all
cluster members and perform local prediction


Cluster member


also perform prediction and transmit
to cluster head only if they are not within a specified
error bound



Prediction model selection


Autoregressive(AR) model




can be calculated by Yule Walker equation or least
square method









p
i
i
t
t
t
x
w
c
x
1
ˆ


Prediction model selection


Lemma
:
Given a linear predictor P, a m
-
step prediction error is less than
m times a one
-
step prediction error .

m
-
step prediction error is .



Given an error bound , the predictor can provide confidence level of







is cummulative distribution of Gaussion white noise


SDV

is the standard deviation


2
2
)]
(
[

m
m
e
VAR


1
)]
(
[
2











m
e
SDV
m



Energy
-
Efficient Algorithm Selection


Tradeoff


Communication cost: Without local prediction, all sensor
nodes will send original data values to the cluster head.


Computation cost: With local prediction.

Energy
-
Efficient Algorithm Selection


Theorem
:
If the error bound satisfies the scheme
with local prediction is more energy
-
efficient


The variance is unknown, use correlation coefficient to
represent m
-
step prediction error




Eliminate , we have
condition (1)
, if error tolerence and
correlation coefficient satisfy




the scheme with local prediction is more energy
-
efficient





k
k
m
2
1














j
x

2
2
1
2
2
)
)
(
1
(
)]
(
[




m
j
m
m
e
VAR
p
j
x
j
x




















k
k
m
j
x
p
j
x
j
2
1
)
(
1
1
1




Algorithm Design
-

cluster head


Pseudo
-
code description







Cluster head will continuously receive data from cluster
members to update the set of history data, or when no data
values are received, will use the predicted value instead of
update.





Algorithm Design
-

cluster head









Periodic process to determine algorithm selection, with or without
local prediction.

Algorithm Design
-

cluster member








Each cluster member maintains a set of history data of its own.


Local prediction disable: transmit the data values


Local prediction enable: perform prediction on each data value, if
not within the error bound, still have to send the value to the cluster
head

Algorithm Design


sleep/wake
scheduling


For some applications may tolerate a few missing value not
within error bound. If confidence level(or having data values
within the error bound) is very high. There is no need for the
nodes to stay awake to obtain data values.

Algorithm Design


sleep/wake
scheduling









Disable local prediction as default. When cluster members is
awake, the cluster head checks if the member
'
s data values are
within the error bound with high probability. If yes, send a
message to power off the member.

Algorithm Design


sleep/wake
scheduling


The condition (2) should be the condition is higher
than ,i.e.






m

threshold



threshold
p
j
x
j
x
j
m

























1
1
2
1
Algorithm Design


sleep/wake
scheduling


To remain accurate prediction, periodic but infrequent collection
from cluster member is still necessary.



Here use a heuristic solution: let be the time interval between
two consecutive report. We set a sleep duration of , when
a member wakes up, it will continuously perform data reading
(and possibly reporting) for the next time.




is initially set to , increased if condition (2) holds, or
decreased if it does not.



*
f
m


*
m
f
m
m
Experiment result


Data
-
set: Intel Berkeley Lab Data


54 nodes spread around the lab


Temperature data within one week


Nodes close to sink set as cluster heads


Nodes report data every 30 seconds



10mJ/Byte, 80mJ per message(assume the value is 8 bytes
double precision number)



13 cluster heads and 40 cluster members, only measure the
consumption of cluster members and LEAP with control of sleep,
set confidence level

9
.
0

threshold

Experiment result










k represents the ratio between transmission/prediction energy
consumption.

(i.e. k=10 and 100, prediction energy consumption is 8 and 0.8
mJ per value)

Conclusion


LEAP is a energy
-
aware data collection
approach


clustering
-
based: sensor nodes form clusters
and cluster head collect and maintain data
values


prediction
-
based: energy
-
aware prediction is
used to find the subtle tradeoff between
communication and prediction cost.