Database Reference
In-Depth Information
form in this particular case is dictated by the hierarchical nature of the
wavelet transform.
A subsequent study [111] generalizes on these ideas, by decoupling the
approximation of the time series from a particular dimension-reduction
algorithm, and employs user-input to specify how the available memory
will be used for the approximation. There has also been relevant work
in machine learning, and more specifically, in the neural network com-
munity, where the main goal is to model time-varying patterns in data
series [10, 29].
A general and ecient solution to the amnesic summarization prob-
lems defined earlier is presented in [70]. This study describes solutions
for the four variations of the problem, based on online algorithms that
use a piecewise linear approximation model. When a new point arrives,
the algorithms update the approximation model in sub-linear time on
the number of linear segments.
It has been shown that the techniques mentioned above can be im-
plemented in a very ecient manner in sensor nodes [89]. Moreover,
amnesic summarization has been studied in the context of flash memo-
ries [67], which offer significant benefits that can be exploited by WSN
deployments.
3. Data Processing
Another interesting and important research direction in the context of
WSN data management is that of ecient data processing and analysis,
and a significant amount of effort has been devoted to it. In this case,
we are interested in supporting different types of complex queries in the
specific, resource-constrained environment of a WSN.
Several frameworks for the ecient execution of queries in a sensor
network have been developed in the past years [60, 59, 103]. The focus
in these works was to propose data processing and optimization meth-
ods geared specifically towards sensor networks, with the early studies
describing in-network aggregation techniques for reducing the amount
of data transmitted by the nodes. Ali et al. [7] propose an interesting
approach to detect and track discrete phenomena (PDT) in sensor net-
works. Hellerstein et al. [49] propose algorithms to partition the sensors
into
isobars
, i.e., groups of neighboring sensors with approximately equal
values during an epoch. Other works have proposed techniques that take
into account missing values, outliers, and intermittent connections [44,
30, 101]. We note that some of the techniques we discussed earlier are
applicable here (e.g., either to answer adhoc queries [31], or
SELECT
*
queries [87]).
