REAL-TIME DATA ANALYTICS IN SENSOR NETWORKS - Managing and Mining Sensor Data - page 184

window window

time PDF time PDF

past future past future

(a) (b)

Figure 7.3. Estimation of data distribution in sliding window for two time instances
(1-d data) [69].

In such an online setting, we require that each sensor maintains a
model for the distribution of values it generates within a sliding window
W (see Figure 7.3 ) . Such a model can be eciently and effectively main-
tained over time. Then, we need to ensure that this mechanism operates
in a distributed fashion. Through a model composition mechanism, we
are able to take the data distribution models of two (or more) streams,
and construct a single model that describes their combined behavior.
The framework also proposes mechanisms for incrementally maintaining
the models across all levels of the (conceptual) hierarchy, as well as for
comparing them in order to determine the similarity of the sensed val-
ues. All the above operations can be eciently supported in real-time
by a sensor node [69].

3.2 Detection and Tracking of Homogeneous
Regions
The first application is identification and tracking of homogeneous
regions [7, 49], which are defined as spatial divisions of the field under
observation that exhibit similar measured values over time, such as an
oil spill detected in the ocean (see Figure 7.4 ). The sensors deployed
around the origin of the spill can organize themselves into a network
and communicate the measurements, to detect regions of varying oil
concentrations.
Recent studies propose methods for delineating homogeneous regions
by a boundary [24, 68]. However, in several situations we need a more
generalized grouping of the sensors, based on the sensed values over a
time interval. In general, we would like to solve the problems of detecting
and tracking such homogeneous regions in real-time when the definition
of the phenomenon is not knowninadvance.
Using the framework described in Section 3.1.1, we can eciently
identify sensors with similar readings, by comparing their models of the
densities of the sensed values [92]. Sensors with very similar models (i.e.,