Database Reference
In-Depth Information
term outlier has many definitions, but the core problem is to detect
points or observations from the dataset that are different compared to
the others. In WSN environments, this translates to finding points which
are compared to all the points that are sensed by the sensors [12]. There
are many applications of anomaly detection in WSNs. Here we briefly
list some of these here:
Environmental monitoring: sensors are deployed in harsh environ-
ments to monitor events that occur in natural environments
Habitat monitoring of species or animals for conversation purposes
or for understanding their migration patterns
Health and medical monitoring tasks in which the goal is to use
different kinds of non-intrusive wearable sensors (
e.g.
acoustic,
temperature, SO
2
, pressure) to analyze the health of humans
Industrial monitoring: sensors are used to sense the health or con-
dition of industrial processes
Target tracking and surveillance, sensors are embedded in moving
targets to track them in real-time
In all of these applications, there is a real need for anomaly detection
for further analysis of abnormal observations. The task of outlier detec-
tion in WSN is extremely dicult mainly because of these reasons [27]:
(1) resource constraints, (2) high communication and computation cost,
(3) distributed streaming data, (4) asynchronous computation model,
(5) large scale deployment, and (6) dynamic network topology, to name
a few. In the remainder of this section we discuss several techniques for
outlier detection in WSNs following the taxonomy given in Chandola
et
al.
[12] and Zhang
et al.
[66].
4.1 Statistical approaches
In statistical approaches, the task is to model the probability distri-
bution of the data using parametric or non-parametric approaches and
then tag as outliers those data points which do not fit the modeled dis-
tribution.
Wu
et al.
[60] present two local techniques for identification of out-
lying sensors. These techniques employ the spatial correlation of the
readings existing among neighboring sensor nodes to detect bad sen-
sors. Each node computes the distance between its own reading and the
median reading of its neighboring sensors. A node is considered as an
outlying node if, the absolute value of this distance is suciently large
