Database Reference
In-Depth Information
Other techniques offer solutions for ecient spatio-temporal data sup-
pression [56, 113, 99, 52, 47, 73], where in addition to the temporal
correlations present in the sensor network data, they aim at identifying
and exploiting the spatial correlations of the data, as well. Furthermore,
previous works have proposed algorithms that help in the selection of
representative nodes when we want to monitor large-scale phenomena
(i.e., phenomena that evolve over days, or months, and involve several
sensor nodes) [6], or when we want to take into account the remaining
energy of each individual node [63]. The above techniques help to fur-
ther reduce the communication cost of the sensor network, and could be
applied on top of the model-driven, or data-driven techniques.
In the rest of this section, we will discuss techniques in the areas
of model-driven and data-driven data acquisition, as well as in spatio-
temporal data suppression.
2.1 Model-Driven Data Acquisition
The aim of the model-driven approach is to (conceptually) collect,
or process queries on all the data sensed by the WSN, based on prob-
abilistic models that capture the correlations that exist in these data.
We note that sensor readings exhibit such correlations in a wide range
of domains and applications. This is true, because often times sensors
are monitoring slow-changing phenomena with high temporal resolu-
tion and/or high spatial resolution. Moreover, correlations may also be
present among different types of readings coming from the same sensor
node (e.g., it has been shown that temperature and voltage readings are
correlated [32]; at the same time it is much less expensive to take voltage
readings than temperature).
The model-driven approach works as follows. During an initial train-
ing phase, all the sensed data are collected from the nodes in the net-
work, in order to train the probabilistic models that are stored in the
sink. Then, these models are used in order to estimate the sensed values,
and additionally provide probabilistic guarantees on the correctness of
these estimates. Therefore, instead of querying the sensors, we operate
on the data produced by the models. If the guarantees produced by the
models for these data do not satisfy the accuracy requirements of the
application, then we can request additional real data values from the
sensors, in order to refine the models to the point that the probabilistic
guarantees satisfy the application requirements.
We can now formally define the model-driven data acquisition prob-
lem.
