Database Reference
In-Depth Information
the agreed behavior. But, unlike other approaches, the definition of
expected behavior proposed in [61] is more generic, and is not limited
toathreshold
δ
.
In this approach a sensor can either be an updater (one who acquires
or forwards sensor values) or an observer (one who receives sensor val-
ues). A sensor node can be both, updater and observer, depending on
whether it is on the boundary of the sensor network or an intermediate
node. The updaters and the observers maintain a model encoding func-
tion
f
enc
and a decoding function
f
dec
. These model encoding/decoding
functions define the agreed behavior of the sensor values. The updater
uses the encoding function to encode the sensor value
v
ij
intoatrans-
mission message
g
ij
, and transmits it to the observer.
The observer, then, uses the decoding function
f
dec
to decode the
message
g
ij
and construct
v
ij
. If the observer finds that
v
ij
has not
changed significantly, as defined by the encoding function, then the ob-
server transmits a
null
symbol. A
null
symbol indicates that the sensor
value is
suppressed
by the observer. Following is an example of the en-
coding and decoding functions [61]:
f
enc
(
v
ij
,v
i
j
)=
g
ij
=
v
ij
−
v
i
j
,
if
|
v
ij
−
v
i
j
|
>δ
;
(2.7)
g
ij
=
null
,
otherwise.
f
dec
(
g
ij
, v
(
i−
1)
j
)=
v
(
i−
1)
j
+
g
ij
,
if
g
ij
=
null
;
(2.8)
v
(
i−
1)
j
,
if
g
ij
=
null
.
In the above example, the encoding function
f
enc
computes the difference
between the model predicted sensor value
v
i
j
and the raw sensor value
v
ij
. Then, this difference is transmitted to the observer only if it is
greater than
δ
, otherwise the
null
symbol is transmitted. The decoding
function
f
dec
decodes the sensor value
v
(
i−
1)
j
using the message
g
ij
.
The encoding and decoding functions in the above example are pur-
posefully chosen to demonstrate how the
δ
threshold approach can be
replicated by these functions. More elaborate definitions of these func-
tions, which are used for encoding complicated behavior, can be found
in [61].
3. Model-Based Sensor Data Cleaning
A well-known characteristic of sensor data is that it is uncertain and
erroneous. This is due to the fact that sensors often operate with dis-
charged batteries, network failures, and imprecision. Other factors, such
as low-cost sensors, freezing or heating of the casing or measurement
device, accumulation of dirt, mechanical failure or vandalism (from hu-