Database Reference
In-Depth Information
In the following paragraphs, we present a framework that enables the
development of a variety of complex processing applications in a sensor
network. These are applications with high processing requirements over
a significant portion of the data generated by the entire WSN.Examples
of such applications are the identification and tracking of homogeneous
regions, and outlier detection. The identification and tracking of ho-
mogeneous regions is used for environmental monitoring (e.g., around
oil-drill, or chemical plant sites). In outlier detection, we are interested
in discovering exceptional situations that may require the attention of a
human analyst: when some of the values of some sensor are not normal,
when the number of abnormal values exceeds a given threshold, or when
the values of a given sensor are significantly different from the values of
its neighbors. We further discuss these applications below.
3.1 Enabling Complex Analytics
The way that streaming applications are able to eciently process
continuous data arriving at high rates, such as those generated byWSNs,
is by computing succinct summaries of the data, and operating on these
summaries [41, 32].
The framework we describe below aims to approximate in an online
fashion multi-dimensional data series distributions [69]. This framework
is adaptive and does not require any a priori knowledge about the dis-
tributions of the sensed values. Moreover, it operates in a distributed
fashion, thus, exploiting all the available resources of the WSN,and
reusing any processing that has already taken place.
3.1.1 Data Distribution Approximation Framework.
The proposed framework for estimating the underlying distribution of
a streaming data series works both for the sliding time window and
the landmark window models [69]. This framework estimates the distri-
bution of the values generated by the sensors using the kernel density
estimators [84], which offer the following desirable properties: (i) they
are ecient to compute and maintain in a streaming environment; (ii)
they can very accurately approximate an unknown data distribution,
with no a priori knowledge and (effectively) no parameters; (iii) they
can easily be combined and (iv) they scale well in multiple dimensions.
The above properties make the framework applicable to large sensor
networks, organized in a hierarchical way
6
[104].
6
The hierarchical decomposition of the sensor network, as well as the selection of the
leaders for each level of the hierarchy, can be achieved using any of the energy-e
cient
techniques proposed in the literature [38, 61, 110].
