Database Reference
In-Depth Information
f(t)
f(t)
f(t)
store
f(t)
outlier
v
v
v
v
value
į
t
t
t
t
t
1
data acquisition
(a)
t
2
t
1
t
2
t
1
t'
t
2
t
1
t
2
data cleaning
(b)
query processing
(c)
data compression
(d)
Figure 2.1.
Various tasks performed by models-based techniques. (a) to improve
acquisitional eciency, a function is fitted to the first three sensor values, and the
remaining values (shown dotted) are not acquired, since they are within a threshold
δ
, (b) data is cleaned by identifying outliers after fitting a linear model, (c) a query
requesting the value at time
t
can be answered using interpolation, (d) only the first
and the last sensor value can be stored as compressed representation of the sensor
values.
Data Acquisition:
Sensor data acquisition is the task responsi-
ble for eciently acquiring samples from the sensors in a sensor
network. The primary objective of the sensor data acquisition
task is to attain energy e
ciency. This objective is driven by
the fact that most sensors are battery-powered and are located in
inaccessible locations (e.g., environmental monitoring sensors are
sometimes located at high altitudes and are surrounded by highly
inaccessible terrains). In the literature, there are two major types
of acquisition approaches: pull-based and push-based. In the pull-
based approach, data is only acquired at a user-defined frequency
of acquisition. On the other hand, in the push-based approach, the
sensors and the base station agree on an expected behavior; sensors
only send data to the base station if the sensor values deviate from
such expected behavior. In this chapter, we cover a representative
collection of model-based sensor data acquisition approaches [2,
12, 17, 16, 18, 27, 28, 41, 66].
Data Cleaning:
The data obtained from the sensors is often er-
roneous. Erroneous sensor values are mainly generated due to the
following reasons: (a) intermittent loss of communication with the
sensor, (b) sensor's battery is discharged, (c) other types of sensor
failures, for example, snow accumulation on the sensor, etc. Model-
based approaches for data cleaning often use a model to infer the
most probable sensor value. Then the raw sensor value is marked
erroneous or outlier if the raw sensor value deviates significantly
from the inferred sensor value. Another important approach for
data cleaning is known as declarative data cleaning [32, 46, 54].
