Database Reference
In-Depth Information
In comparison to hierarchical control clustering, DWEHC converges
in a constant number of iterations. Moreover, it not only respects the
distance between nodes, but also their residual energy. In contrast to
previously proposed protocols like LEACH [33], DWEHC doesn't require
knowledge about the network size, density or homogeneity or about the
number of levels, like HEED [63]. While cluster topologies generated by
HEED may not achieve minimum energy consumption in intra-cluster
communication, it was shown empirically that the energy savings of
DWEHC outperform those or HEED. Also, DWEHC produces more
well-balanced clusters and a better distribution of cluster heads, result-
ing in higher energy savings for inter-cluster communication.
2.1.3 Further Reading.
Hierarchical node clustering and
DWEHC are only representatives of several distributed clustering al-
gorithms that have been developed for WSNs. The survey article by
Abbasi and Younis [1] gives a thorough summary of many additional
algorithms. For example, other clustering approaches that have a linear
convergence rate are LCA [4], CLUBS [42], RCC [42] and EEHC [5]. Fur-
ther approaches with a constant number of iterations are, for example,
LEACH [33], HEED [63], MOCA [64], EECPL [3] or N-LEACH [56].
2.2 Distributed Clustering of Sensor
Measurements
The distributed algorithms described in the previous section cluster
sensor nodes. Their purpose is to determine a node topology that allows
for an energy-e
cient gathering of data,
i.e.
sensor measurements, from
the network. As an unsupervised method, clustering can also be used
for an exploratory analysis of data, finding groups of similar sensor mea-
surements. Traditional clustering algorithms usually assume all data to
be available at a single site, like a base station. Even with an established
network topology that allows for energy-ecient communication, due to
energy-constraints it is usually not feasible to transfer all available sen-
sor measurements to a single site for clustering. Instead, distributed
algorithms need to process data in-network, locally at the sensor nodes,
and respect the given limitations of WSNs as much as possible when
communicating with other nodes.
Distributed clustering algorithms have been developed in distributed
data mining (DDM). These algorithms are often based on the parallel
computing paradigm. Running time should be improved by moving data
over high-bandwidth connections from a central location to so called
compute nodes, and then working on subsets of the data in parallel.
