Database Reference
In-Depth Information
data. They also assume an unlimited amount of available energy. There-
fore, the algorithms won't work in the highly constrained setting of dis-
tributed wireless sensor nodes. For WSNs, the algorithms have been
either modified or new clustering algorithms have been developed that
take into account the distributed nature of sensors and the severe com-
munication constraints due to limited battery power and bandwidth. As
will become apparent, in WSNs usually multiple quality criterions can be
applied, turning clustering into a multi-objective optimization problem.
In the next section it is shown that, while often not directly focused
on data analysis, clustering algorithms play a crucial rule in creating
communication ecient topologies of sensor nodes. Especially, the solu-
tions found for grouping sensor nodes may inspire the design of future
energy-constrained data analysis methods. The section afterwards then
discusses already existing distributed clustering algorithms for data anal-
ysis,
i.e.
of sensor measurements.
2.1 Distributed Clustering of Sensor Nodes
Continuous monitoring as well as intermittent querying of sensor net-
works involves transmitting data from individual sensor nodes, the
sour-
ces
, to a single node, the
sink
. Communication costs increase with higher
distance
r
between sensor nodes, as ground reflections from short an-
tenna heights may cause a drop-off of the radio signal power by
r
4
[47].
Therefore, hierarchical, tiered
multi-hop
architectures with shorter dis-
tances between relaying nodes are usually more energy-ecient than
letting all sensors communicate directly with some base station [25].
The sensor nodes in tiered multi-hop networks form — possibly hi-
erarchical — clusters and certain nodes in each cluster are designated
as
cluster heads
. Cluster heads fulfill special roles, like relaying signals
from local nodes in their cluster to other cluster heads or a base station.
They also can manage and restrict network access as well as the life cycle
of local nodes, or reduce the amount of data transmitted by aggregating
and pre-processing the signals from sensor nodes in their cluster.
Manual placement of sensors and routing through pre-determined
paths are only feasible for very small networks. However, typical ap-
plications of sensor networks, like environmental monitoring, disaster
management or military surveillance missions envision hundreds or even
thousands of sensor nodes [1], possibly deployed randomly,
e.g.
dropped
by a helicopter. The network is usually left unattended for long periods
of time and batteries can't be recharged. While some setups utilize mo-
bile sensors, sensor nodes are usually assumed to operate stationary af-
ter deployment. Nevertheless, the network could change over time, since
