Database Reference
In-Depth Information
that is available about the distribution of the data series and its errors.
PROUD requires to know the standard deviation of the uncertainty er-
ror and a single observed value for each timestamp, and assumes that
the standard deviation of the uncertainty error remains constant across
all timestamps. DUST takes as input a single observed value of the data
series for each timestamp, and in addition, needs to know the distri-
bution of the uncertainty error for each single time stamp, as well as
the distribution of the values of the data series. This means that, in
contrast to PROUD, DUST can take into account mixed distributions
for the uncertainty errors (albeit, they have to be explicitly provided
in the input). MUNICH does not need to know the distribution of the
data series values, or the distribution of the value errors: it simply oper-
ates on the observations available at each timestamp. When we do not
have enough, or accurate information on the distribution of the errors,
PROUD and DUST do not offer an advantage in terms of accuracy when
compared to Euclidean [27].
All three techniques are based on the simplifying assumption that
the values of the data series are independent from one another, which
is not true for WSN measurements. A recent study [28] demonstrates
that removing this assumption is beneficial: it proposes the UMA and
UEMA filters (based on the weighted moving average technique), that
in combination with Euclidean distance lead to more accurate results.
These results suggest that more work is needed on techniques that take
into account the temporal correlations that exist in data series.
The time complexity of these techniques is another important factor.
We note that MUNICH is applicable only in the cases where the standard
deviation of the error is relatively small, and the length of the data
series is also small (otherwise the computational cost is prohibitive),
which makes this technique applicable in cases where the sink can do
the processing. To a (much) lesser extent, this is also true for PROUD
and DUST. On the other hand, UMA and UEMA have significantly
lower resource requirements, and could be eciently implemented in a
sensor node.
4.3 Ubiquitous Sensor Networks
Lots of work and research effort has been devoted in the past years
in the study of various problems related to WSNs. Several ecient
techniques have been developed for the acquisition, management, pro-
cessing, and analysis of the sensed data, and at the same time (different
forms of) WSNs are being deployed in increasingly more domains and
situations.
