Databases Reference
In-Depth Information
tensive retraining and a huge number of training data sets for effective general-
ization. Furthermore, the centralized processing or single-processing approach
used in existing schemes creates some significant problems. For example, the
constant flow of sensory data results in high communication overheads; re-
routing procedures and relocation activities of sensor nodes that often occur
in real-time applications; and significantly long delays in detecting critical
events when computational bottlenecks are present. These problems limit the
event detection schemes' ability to scale up to massive sensory data processing.
Artificial neural networks (ANNs) and other machine learning techniques
are the most commonly applied classification techniques in event detection
schemes for WSNs. Some of these schemes implement the Kohonen Self-
Organizing Map (SOM) or other activation-based neural networks, such as
the Radial Basis Function (RBF) neural network. Because of their learning
complexity and highly cohesive training-validation approach, these schemes
cannot scale up to the dynamics of the WSN network.
9.1.1 WSN Event Detection
Breakthroughs in communication technologies have enhanced the perfor-
mance of existing coarse-grained networks, such as cloud and grid computing.
Research has also led to the rapid growth of emerging fine-grained networks,
such as wireless sensor networks (WSNs). These networks emerged from the
confluence of wireless communication, extensive computational schemes, and
sophisticated sensor technology. For example, WSNs are created from a collec-
tion of self-organized wireless and battery-powered devices that have sensing
capabilities. The emergence of these self-organized networks of tiny processing
devices has led to the ability for parallel and distributed computing deploy-
ment in fine-grained systems.
Unfortunately, the current scenario in WSN deployment is still far away
from its tremendous potential. A WSN has only been demonstrated for hum-
ble applications such as meter reading in buildings and a basic form of eco-
logical monitoring. Achieving the full potential of this technology requires the
development of an intelligent computational scheme.
Common approaches implemented in existing WSN applications usually
involve a number of processing steps, including sensory data capture and con-
veyance of these data to a central entity, known as the base station, for further
refinement and analysis. Consequently, if it is called up for widespread use,
this approach leads to a system bottleneck. Because of latency between the
data capture/aggregation and processing time, processing delays occur inter-
mittently. These limitations make WSNs less suitable for real-time monitoring
applications. Therefore, a new approach for data processing in a WSN must
be designed that has the ability to process the sensory data in situ and in
a decentralized manner and can generate highly condensed and sophisticated
outputs internally. These abilities alleviate the bottleneck problem in WSNs
