Database Reference
In-Depth Information
Another idea that has been studied is to select a set of representative
nodes, and use only those for transmitting measurements to the sink.
The premise is that each representative node has measurements similar
to the measurements of the nodes in its neighborhood. Then, it is only
the representative nodes that need to communicate the sensed values to
the sink, thus, significantly reducing the energy spent by the WSN.
Data mining approaches contributed to this problem, by providing
techniques for clustering and selecting representatives [46, 80, 62, 108].
Inside each cluster, the node with the most similar readings to the mea-
surements of all nodes inside that cluster is selected as a cluster rep-
resentative. Many algorithms were developed to deal with the online
distributed clustering of data.
SERENE [9] is a framework for SElecting REpresentatives in a sensor
NEtwork. It uses clustering techniques to select the subset of nodes that
can best represent the rest of sensors in the network. In order to select an
appropriate set of representative sensors, SERENE performs an analysis
of the historical readings of sensor nodes, identifies the spatio-temporal
correlations among sensors (based on their readings), and groups sen-
sors into clusters according to these correlations. Then, each cluster
performs further analysis in order to select the sensors with the highest
representation quality. We note that the analysis of the historical data,
which has to be repeated when the distribution of the sensor readings
changes, may take place in the sensors or in the sink, according to the
amount of resources required.
Snapshot Queries [56] is another approach that introduces a platform
for energy e
cient data collection in sensor networks. By selecting a
small set of representative nodes, this approach provides responses to
user queries and reduces the energy consumption in the network. In
order to select representatives, each sensor node in this approach builds
a data model of the distribution of measurement values of its neighbors
for each attribute. After a node decides which of its neighbors it can
effectively represent, it broadcasts its list of candidate cluster members
to all its neighbors. Each node selects as its representative the neigh-
bor that can represent it, and that additionally has the longest list of
candidate cluster members.
In ECLUN [47], nodes do not continuously communicate with the rep-
resentatives, but communication is established only when a state change
is detected in the monitored phenomena. This communication is further
reduced through the careful construction of clusters, which considers
similarity in sub-spaces of the full-dimensional sensor readings space.
This makes the above approach suitable to deployments of sensor node
that produce multi-dimensional readings (i.e., monitor several phenom-
