Distributed Event Identification for WSNs in

Non-Stationary Environments

Kamran Ali*, S.B. AliK, I. H. Naqvi, M. A. Lodhio

Department of Electrical Engineering, LUMS School of Science and Engineering, Pakistan
*Department of Computer Science and Engineering, Michigan State University, USA
KTU Munchen, Germany

oDepartment of Electrical and Computer Engineering, Rutgers, USA

1
 Outline
Introduction
- Outlier, Event Detection & Event Identification in WSNs
Online & Adaptive Clustering for Outlier
detection
Event Detection
Event Identification Simulation and
Experimental Results

2
 Introduction
The data obtained from WSN can be classified as:
- Normal
- Outliers-Anomalous data due to noise or sensor faults
- Events
The purpose is to separate outliers & events and
determine the attributes involved in the event (Event
Identification).
Characteristics of detection and identification techniques
in Harsh and Non-stationary environments:
- Unsupervised, online, distributed and adaptive
- Multivariate data
- Spatio-Temporal and attribute correlations
- Insusceptible to non-stationarity & dynamic network
topology
3
 Outlier Detection using Online and
Adaptive Clustering
Consider a WSN with strong temporal, attribute &
spatial correlations, where each node collects a total of
k, d dimensional data samples Xk.
During the outlier detection phase, the hyper-ellipsoid
clustering algorithm encapsulates the data samples by
using the following equations:

4
 Outlier Detection and Non stationary
environments
For a newly arrived data sample, the mean and the second
moment can be updated in the following manner:

For non stationary environments, we make use of independent

forgetting factor approach; Where is called factor, used for
environment adaptation:

the updated mean at the kth iteration can be written as:

5
Event Detection

6
leave one out (LOO) event identification
algorithm
Step1: Projection of the hyper-ellipsoid
(containing the data set eX i.e. EQ, its covariance
SE and mean mE) on to a subspace Xq which
contains all but one (qth) attributes
Correlation of all (q-1) attributes is taken into
account
Step2: Finding the Mahalanobis distance

Step3: Decision Formulation

7
Synthetic dataset for Simulations

A 4-dimensional synthetic data set with correlated and uncorrelated events

8
Simulation Results on Synthetic dataset

9
Performance of clustering techniques:
synthetic data set

A 4-dimensional synthetic data set with correlated and uncorrelated events 10

 Real Data sets
We tested our algorithms on three real data sets
Multi-hop WSN deployment using TelosB motes
measuring humidity and temperature at 5 second
intervals [18]
Data from epochs 25000 to 30000 of node 18 of the
IBRL (Intel Berkeley Research Lab) dataset [19]
Data collected at station 10 of the network deployed
on the rock glacier in Switzerland in a 12 days
collection period starting from October 10th 2007
[20]

11
12
Event detection and identification
results: real data sets

13
 Hardware Implementation and Results
Two CC430 based WizziMotes
running DASH7 protocol were used.
Temperature and Illumination
attributes were tested.
Tested using white LED and
fluorescent lights. WizziBase with
WizziMote
- Scenario 1 (turning off the lights of the room in which the
motes were deployed) 0.5% and 99.5% contribution for
temperature and illumination respectively were reported.

- Scenario 2 (illuminated two incandescent bulbs in a close

vicinity of both the motes) 76.34% and 23.66% contribution
for temperature and illumination respectively were reported.
14
 Conclusion and Future work
Proposed scheme gives accurate identification
ratios in cases where both independent as well as
correlated attributes contribute towards an
event.
Future work entails:
Implementation in a larger WSN framework
More thorough evaluations of our proposed
schemes in terms of required memory resources and
the overall network lifetime.

15
Thanks and Questions???

16