Sei sulla pagina 1di 34

Outline

Introduction and Related Works


Proposed Method
Simulation
Conclusion

.
.
Reducing The Data Transmission in Wireless
Sensor Networks Using The Principal Component
Analysis
.
.. .

.
A. Rooshenas, H. R. Rabiee, A. Movaghar, M. Y. Naderi

The Sixth International Conference on Intelligent Sensors,


Sensor Networks and Information Processing
ISSNIP 2010

8 Dec 2010
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 1 / 34 ’
Outline
Introduction and Related Works
Proposed Method
Simulation
Conclusion

.
. .1 Introduction and Related Works
Background
PCAg
DPCA
.
2 . Proposed Method
.
Using PCA
LocalPCA: Using Collective PCA
Reconstruction Error
.3. Simulation
.
Settings
Results
.4. Conclusion
.
Future Works
References
Appendix
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 2 / 34 ’
Outline
Introduction and Related Works Background
Proposed Method PCAg
Simulation DPCA
Conclusion

. Problem Definition

.
Different Power Consumption in Routing Tree .
..
Leaf Nodes
Intermediate Node
.
.. .

.
.
Problems .
..
Intermediate nodes should relay other packets to the base
station.
Lifetime of intermediate nodes affect the longevity of Wireless
Sensor Networks
.
.. .

.
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 3 / 34 ’
Outline
Introduction and Related Works Background
Proposed Method PCAg
Simulation DPCA
Conclusion

. Dimension Reduction

.x2

.v

. .x1

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 4 / 34 ’
Outline
Introduction and Related Works Background
Proposed Method PCAg
Simulation DPCA
Conclusion

. Dimension Reduction

.x2

.v
.y = vT x

.v

. .x1
.

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 5 / 34 ’
Outline
Introduction and Related Works Background
Proposed Method PCAg
Simulation DPCA
Conclusion

. Dimension Reduction

.x2

.v
.y = vT x

.v

.x̂ = vy
. .x1
.

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 6 / 34 ’
Outline
Introduction and Related Works Background
Proposed Method PCAg
Simulation DPCA
Conclusion

. Related Works

Principal Component Aggregation (PCAg).


Unsupervised
Supervised
Distributed Principal Component Analysis (DPCA).

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 7 / 34 ’
Outline
Introduction and Related Works Background
Proposed Method PCAg
Simulation DPCA
Conclusion

. Principal Component Aggregation

.BS

.x1 , x2 , x3 , x4 , x5 .x6 , x7 , x8 , x9

.n3 .n6 .n9


.x1 , x2 , x3 .x7 , x8 , x9
Base station gathers
observations from sensor .x1 , x2 .x4 , x5 .x7 , x8

networks .n2 .n5 .n8

.x1 .x4 .x7

.n1 .n4 .n7

. . . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 8 / 34 ’
Outline
Introduction and Related Works Background
Proposed Method PCAg
Simulation DPCA
Conclusion

. Principal Component Aggregation

.BS

.v1 , v2 , v3 , v4 .v5 , v6 , v7 , v8 , v9

.n3 .n6 .n9


.v1 , v2 , v3 .v7 , v8 , v9
Base station calculates
eigenvectors and sends .v1 , v2 .v4 , v5 .v7 , v8

them back to network .n2 .n5 .n8

.v1 .v4 .v7

.n1 .n4 .n7

. . . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 9 / 34 ’
Outline
Introduction and Related Works Background
Proposed Method PCAg
Simulation DPCA
Conclusion

. Principal Component Aggregation

Sensor nodes send partial .BS

PCs in each epoch ∑9


.y = i=1 vi ∗ xi
∑3
The base station . i=1 vi ∗ xi
.n3 .n6 ∑ .n9
reconstructs observations . 9i=7 vi ∗ xi
∑2 ∑5 ∑8
Reconstruction error is not . i=1 vi ∗ xi . i=4 vi ∗ xi . i=7 vi ∗ xi

available at the base .n2 .n5 .n8


station
.v1 ∗ x1 .v4 ∗ x4 .v7 ∗ x7
The base station calculates
eigenvectors periodically .n1 .n4 .n7

. . . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 10 / 34 ’
Outline
Introduction and Related Works Background
Proposed Method PCAg
Simulation DPCA
Conclusion

. Principal Component Aggregation

The base station sends PCs to the .BS


network .y
Each sensor can calculate its
.n3 .n6 .n9
reconstruction error .y .y

ei = ∥xi − y ∗ vi ∥ .y .y .y

Updating one observation may .n2 .n5 .n8


change all other elements of an
.y .y .y
eigenvector
Update decision is ambiguous .n1 .n4 .n7

. . . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 11 / 34 ’
Outline
Introduction and Related Works Background
Proposed Method PCAg
Simulation DPCA
Conclusion

. Distributed Principal Component Analysis

Assumptions:
Each sensor receives sensed value from its neighbors
Only the data of sensors which are in the radio range of each
other are correlated
For the ∑
computation of eigenvectors, DPCA uses:
vit+1 = N j=1 Cij ∗ vj
t

Each sensor broadcasts its eigenvector’s entry vi and its


sensed value.
Assumptions of DCPA are not supported by weather data

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 12 / 34 ’
Outline
Introduction and Related Works Using PCA
Proposed Method LocalPCA: Using Collective PCA
Simulation Reconstruction Error
Conclusion

. Using PCA for Data Reduction

.
Retained Variance for 50 sensors .
..
Retained Variance λ1 /λ2
Temperature 1.0 3.6909e+015
Humidity 1.0 3.4407e+015
Light 1.0 2.4402e+015
Voltage 1.0 3.1737e+015
Combined 1.0 1.9305e+015
.
.. .

.
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 13 / 34 ’
Outline
Introduction and Related Works Using PCA
Proposed Method LocalPCA: Using Collective PCA
Simulation Reconstruction Error
Conclusion

Leaf nodes send their


observations
Intermediate nodes .BS .
.x̂ = gy
compute PCs
.y = f (x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 )
Intermediate nodes
recalculate PCs according .n8
to reconstruction error
.y1 = f (x1 , x2 , x5 ) .x6 .y2 = f (x3 , x4 , x7 )
The next level
intermediate nodes .n5 .n6 .n7
reconstruct observation
.x1 .x2 .x3 .x4
and compute PCs from all
.n1 .n2 .n3 .n4
downstream observations
The base station
reconstructs observations
. . . . . .

DML 8 Dec 2010 Amirmohammad


. Rooshenas Reducing The Data Trans. in WSN using PCA 14 / 34 ’
Outline
Introduction and Related Works Using PCA
Proposed Method LocalPCA: Using Collective PCA
Simulation Reconstruction Error
Conclusion

. Updating Eigenvectors
Eigenvectors are common between
each pair of parent-child
intermediate nodes .g = g8
.
.BS .
Nodes which are in the vicinity of
.g8 .
the base station must send large
.n8
eigenvectors.
. .g5 .g7 .
For these nodes, the computation .
of the covariance matrix and .n5 .n7
.n6
eigenvectors is energy consuming . . . .

Updated eigenvectors must be sent .n1 .n2 .n3 .n4


to the base station
The base station calculates final
.
eigenvectors . . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 15 / 34 ’
Outline
Introduction and Related Works Using PCA
Proposed Method LocalPCA: Using Collective PCA
Simulation Reconstruction Error
Conclusion

. Features of Using PCA

.
Features of Using PCA .
..
Each intermediate node has error threshold that can be
configured by the base station
High threshold value causes low accuracy and high efficiency
An intermediate node only sends PCs in each epoch
.
.. .

.
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 16 / 34 ’
Outline
Introduction and Related Works Using PCA
Proposed Method LocalPCA: Using Collective PCA
Simulation Reconstruction Error
Conclusion

. Collective PCA

Reconstruction of data in each intermediate node and


combination again is not scalable
Collective PCA: PCA is invariant to an orthonormal linear
transformation
We want to find the principal component of x = [xT T
1 , x2 , s]
T

where s is sensed value.


y1 and y2 are principal compenents of x1 and x2
 
y1
z = y2  and g is the eigenvector of zzT then

s
the principal component of x is y = gT z
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 17 / 34 ’
Outline
Introduction and Related Works Using PCA
Proposed Method LocalPCA: Using Collective PCA
Simulation Reconstruction Error
Conclusion

. Aggregation of PCs using Collective PCA


Leaf nodes send their
observations
.x̂ = gy
Intermediate nodes .BS .
compute PCs
.y = f (y1 , y2 , x6 , x8 )
The next level
intermediate nodes .n8
compute PCs from
.y1 = f (x1 , x2 , x5 ) .x6 .y2 = f (x3 , x4 , x7 )
observations and received
PCs .n5 .n6 .n7

The base btation .x1 .x2 .x3 .x4


reconstructs observations .n1 .n2 .n3 .n4

Intermediate nodes only


send PCs in each epoch . . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 18 / 34 ’
Outline
Introduction and Related Works Using PCA
Proposed Method LocalPCA: Using Collective PCA
Simulation Reconstruction Error
Conclusion

. Updating Eigenvectors
Each intermediate node
.g = [g8 (1) ∗ g5T , g8 (2) ∗ g7T , g8 (3), g8 (4)]T
has its eigenvectors .BS .
Each intermediate node
.g5 , g7 , g8 .
updates its eigenvectors
independently .n8
Updated eigenvectors must
. .g5 . .g7 .
be sent to the base station
The base station calculates .n5 .n6 .n7
final eigenvectors . . . .
Update rate depends on .n1 .n2 .n3 .n4
the number of changes in
environment
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 19 / 34 ’
Outline
Introduction and Related Works Using PCA
Proposed Method LocalPCA: Using Collective PCA
Simulation Reconstruction Error
Conclusion

Proposed method: An aggregation service for calculation


. of reconstruction Error

In PCAg, the base station needs to update eigenvectors in


order to compute reconstruction error
In LocalPCA, the base station needs to tune thresholds
In PCAg and LocalPCA, the base station do not access the
real observation to calculate reconstruction error
We proposed an aggregation service facilitating the
calculation of reconstruction error at the base station.
EAPCAg : We enhance PCAg using this aggregation service

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 20 / 34 ’
Outline
Introduction and Related Works Using PCA
Proposed Method LocalPCA: Using Collective PCA
Simulation Reconstruction Error
Conclusion

Proposed method: An aggregation Service for Calculation


. of Reconstruction Error


x̂ − x 2 = (x̂ − x)T (x̂ − x)
= x̂T x̂ − x̂T x − xT x̂ + xT x

| {zA} A x − 2x AA x + x x
= xT A A T T T T T

I
= x x − xT AAT x
T

= xT x − y T y

n ∑
p
= xi2 − yi2 p≪n
i=1 i=1
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 21 / 34 ’
Outline
Introduction and Related Works
Settings
Proposed Method
Results
Simulation
Conclusion

. Data Set
.
7500 epoch of 10 sensors from Intel-lab .
..

.
.. .

.
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 22 / 34 ’
Outline
Introduction and Related Works
Settings
Proposed Method
Results
Simulation
Conclusion

. Simulation Senario

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 23 / 34 ’
Outline
Introduction and Related Works
Settings
Proposed Method
Results
Simulation
Conclusion

. Packet
.
Packet Size .
..
Field Size
Header 10
Observation 2
Eigenvector 4
element
Principal 4
Component
Squared 4
observation
Request for 1
. observation
.. .

.
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 24 / 34 ’
Outline
Introduction and Related Works
Settings
Proposed Method
Results
Simulation
Conclusion

. Numerical Results

Numerical Results
Transmitted Error Error Error Parameters
Bytes Mean Variance Max
PCAg 200340 0.82 0.96 8.34 update: 11 epoch
LocalPCA 168114 0.82 0.07 1.43 threshold at 23:1.0
threshold at 22,21: 0.35

EAPCAg 215216 0.82 0.23 6.16 threshold:1.42

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 25 / 34 ’
Outline
Introduction and Related Works
Settings
Proposed Method
Results
Simulation
Conclusion

. Discussion

For the sensor node 23rd, EAPCAg method has sent about
15K bytes more than PCAg algorithm
PCAg method has sent about 32K bytes more than LocalPCA.
The average updating interval in EAPCAg is about 13.84
epochs which is around 3 epochs more than PCAg.
EAPCAg has less efficiency than PCAg because the size of
each PC packet is increased from 14 bytes to 18 bytes
In PCAg and EAPCAg methods, when the number of sensor
nodes increases, the eigenvector packets should be divided to
smaller packets, so the amount of transmitted data will
increase
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 26 / 34 ’
Outline
Introduction and Related Works
Settings
Proposed Method
Results
Simulation
Conclusion

.
Comparision of LocalPCA with PCAg at the Base Station .
..

.
.. .

.
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 27 / 34 ’
Outline
Introduction and Related Works
Settings
Proposed Method
Results
Simulation
Conclusion

.
Error of LocalPCA .
..

.
.. .

.
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 28 / 34 ’
Outline
Introduction and Related Works
Settings
Proposed Method
Results
Simulation
Conclusion

. .
Error of PCAg . Error of EAPCAg
. ..

. .
.. . .. .
.

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 29 / 34 ’
Outline
Introduction and Related Works Future Works
Proposed Method References
Simulation Appendix
Conclusion

. Future Works

.
PCA .
..
Efficiency of LocalPCA depends on thresholds
Thresholds should be adjusted adaptively
Weather data has low dimensions. LocalPCA will show better
performance with high dimensional data (e.g. image )
Kernel PCA is an non-linear version of PCA and can handle
data which changes rapidly
.
.. .

.
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 30 / 34 ’
Outline
Introduction and Related Works Future Works
Proposed Method References
Simulation Appendix
Conclusion

Borgne, Y. L., Raybaud, S., and Bontempi, G. Distributed


principal component analysis for wireless sensor networks.
Sensors 8 (2008)
Hillol, K., Huang, W., Krishnamoorthy, S., Byung-Hoon, P.,
and Wang, S. Collective principal component analysis from
distributed, heterogeneous data. Lecture Notes in Computer
Science. 2000.
Jollife, I. Principal Component Analysis, second ed. Springer,
2002.
Y. L. Borgne and G. Bontempi. Unsupervised and supervised
compression with principal component analysis in wireless
sensor networks. in Proc. First International Workshop on
Knowledge Discovery from Sensor Data, 13th ACM SIGKDD
International Conference on Knowledge Discovery and Data
Mining, 2007
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 31 / 34 ’
Outline
Introduction and Related Works Future Works
Proposed Method References
Simulation Appendix
Conclusion

.
Questions? .
..

.
.. .

.
. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 32 / 34 ’
Outline
Introduction and Related Works Future Works
Proposed Method References
Simulation Appendix
Conclusion

. How to find eigenvectors

Computation of Covariance Matrix


Σ = E [(x − E [x])(x − E [x])T ]
Solve eigenvector, eigenvalue equation Σα = λα
if λi is largest eigenvalue and Σαi = λi αi
then αi is dominant eigenvector.
When the dimension reduction from n dimension to q
dimension is efficient? ∑ q
λk
Retained Variance: H(q) = ∑k=1
n
k=1 λk

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 33 / 34 ’
Outline
Introduction and Related Works Future Works
Proposed Method References
Simulation Appendix
Conclusion

. Power Iteration Method

Calculation of eigenvectors by eigenvector, eigenvalue


equations is time consuming O(n3 ).
Power Iteration Method approaches to dominant eigenvector
very fast O(n2 ).
PIM iteration: gk+1 = Σgk
FastPCA : Extension of PIM for calculation of other
eigenvectors.

. . . . . .

DML 8 Dec 2010 Amirmohammad Rooshenas Reducing The Data Trans. in WSN using PCA 34 / 34 ’

Potrebbero piacerti anche