Database Reference
In-Depth Information
the rooms that are
Hot
with probability greater than 0
.
9, then the in-
network model index can eciently prune the rooms that are surely not
a part of the query response. Due to the lack of space, we shall not
cover the details of index construction and pruning. We encourage the
interested reader to read the following paper [5].
4.6 Processing Event Queries
Event queries are another important class of queries that are proposed
in the literature. These queries continuously monitor for a particular
event that could probably occur in sensor data. Consider a setup con-
sisting of RFID sensors in a building. An event query could monitor
an event of a person entering a room or taking coffee, etc. Moreover,
event queries can also be registered, not only to monitor a single event,
but a sequence of events that are important to the user. Again, due to
space constraints, we shall not cover any of the event query processing
approaches in detail. The interested reader is referred to the prior works
on this subject [55, 65, 68, 45].
5. Model-Based Sensor Data Compression
Recent advances in sensor technology has resulted in the availability
of a multitude of (often privately-held) sensors. Embedded sensing func-
tionality (e.g., sound, accelerometer, temperature, GPS, RFID, etc.) is
now included in mobile devices, like, phones, cars, or buses. The large
number of these devices and the huge volume of raw monitored data
pose new challenges for sustainable storage and e
cient retrieval of the
sensor data streams. To this end, a multitude of model-based regression,
transformation and filtering techniques have been proposed for approxi-
mation of sensor data streams. This section categorizes and reviews the
most important model-based approaches towards compression of sensor
data. These models often exploit spatio-temporal correlations within
data streams to compress the data within a certain error norm; this is
also known as
lossy compression
. Moreover, several standard orthogonal
transformation methods (like, Fourier or wavelet transform) reduce the
amount of storage space required by reducing the dimensionality of data.
Unlike the assumptions of Section 2, where we assumed a sensor net-
work consisting of several sensors, here we assume that we only have
a single sensor. We have dropped the several sensors assumption to
simplify the notation and discussion in this section. Furthermore, we
assume that the sensor values from the single sensor are in a form of a
data stream
. Let us denote such a data stream as a sequence of data
tuples (
t
i
,v
i
), where
v
i
is the sensor value at time
t
i
.
