Data mining in wireless sensor networks based on artificial neural-networks algorithms

1
Data mining in wireless sensor networks based on
artificial neural-networks algorithms

• Authors Andrea Kulakov and Danco Davcev
• Presentation by Niyati Parikh

2
Motivation

• Centralized data clustering in sensor networks is
  difficult, not scalable, limited communication
  bandwidth, limited power supply, data redundancy
• Advantage of Neural Networks demand of
  compressed summaries of large spatio-temporal
  data, similarity queries finding similar
  patterns or detecting correlations
• Unsupervised learning ANN perform dimensionality
  reduction or pattern clustering

3
Adaptive Resonance Theory(ART1)
Attentional subsystem
F2
Category layer
reset
Orienting subsystem
-
F1
p
Comparison layer

F0
Input layer
Binary input
4
Adaptive Resonance Theory(ART1)
Attentional subsystem
F2
Category layer
Ti wi . x
reset
B wi
Orienting subsystem
-
F1
p
Comparison layer

F0
Input layer
Binary input
5
Adaptive Resonance Theory(ART1)
Attentional subsystem
F2
Category layer
Ti wi . x
reset
B wi
Orienting subsystem
-
F1
p
Comparison layer
wi . x

x
F0
Input layer
Binary input
6
Adaptive Resonance Theory(ART1)
Attentional subsystem
F2
Category layer
Winew
Ti wi . x
reset
B wi
Orienting subsystem
-
F1
p
Comparison layer
wi . x

x
F0
Input layer
Binary input
7
ART1

• Continue finding an F2 node until prototype
  matches the input well enough or else allocate a
  new F2 node
• Capable of refining learned categories and
  finding new patterns
• Value of p higher the vigilance level, more
  specific clusters

8
FuzzyART

• Same as ART1, but replace intersection operator
  of ART1 with fuzzy set theory conjunction MIN
  operator
• ART1 and FuzzyART use complement coding
  concatenate input pattern b with b or bi with
  (1-bi)
• Look at the features consistently present or
  absent from a pattern

9
Proposed architectures of sensor networks
Clusterhead collecting all sensor data from its
cluster of units
10
One clusterhead collecting and classifying the
data after they are once classified at the lower
level
11
Results
p 0.70 0.85 0.9 0.93 0.95 0.97 0.98 0.99
categories 2 3 8 19 36 87 151 370
12
Comparison

• Tested data robustness made one sensor
  defective
• Architecture1 trained with p0.93 and tested
  with p 0.90
• Architecture2 trained with p0.80 and tested
  with p 0.70
• Architecture2 makes 0.75 classification error

13
Future work

• Applying ARTMAP and FuzzyARTMAP
• - supervised learning versions