MINING SENSOR DATA STREAMS - Managing and Mining Sensor Data - page 144

2.1 Data Uncertainty and Volume
Typical sensor mining applications collect large amounts of data,
which is often subject to uncertainty and errors. This is because sensors
often have errors in data collection and transmission. In many cases,
where the battery runs out, the data may also be incomplete. There-
fore, methods are required to store and process the uncertainty in the
underlying data. A common technique is to perform model driven data
acquisition [36], which explicitly models the uncertainty during the ac-
quisition process. Furthermore, new methods must be constructed in
order to explicitly use these uncertainty models during data processing.
A detailed discussion of a variety of methods for modeling and mining
uncertain data are proposed in [14].
A variety of techniques may be used in order to handle the large vol-
ume of sensor data. One method may be to simply lower the sampling
rate for capturing or transmitting the data. This reduces the granularity
of the data collection process. Many techniques have also been proposed
[34, 35] in order to reduce or compress the volume of the data in the sen-
sor network. Another approach is to discard parts of the data even after
collection and transmission. For example, when the incoming rate of
the data streams is higher than that can be processed by the system,
techniques are required in order to selectively pick data points from the
stream, without losing accuracy. This technique is known as loadshed-
ding . Since sensor generated data streams are generated by processes
which are extraneous to the stream processing application, it is not pos-
sible to control the incoming stream rate. As a result, it is necessary
for the system to have the ability to quickly adjust to varying incoming
stream processing rates. One particular type of adaptivity is the ability
to gracefully degrade performance via “load shedding” (dropping unpro-
cessed tuples to reduce system load) when the demands placed on the
system cannot be met in full given available resources. The loadshedding
can be tailored to specific kinds of applications such as query process-
ing or data mining. A discussion of several loadshedding techniques are
provided in [4]. Finally, a method for reducing the data volume is that
of sensor selection in which data from only a particular set of sensors is
transmitted at a particular time, so that most of the essential informa-
tion is retained. This is also useful for reducing the power requirements
of transmission. We will discuss this method in the next subsection.