Database Reference
In-Depth Information
Online event detection.
In the online case, sensor networks
reduce the bandwidth requirements of data collection by suppressing re-
sults that conform to the model or compressing the data stream through
a model representation. This has coincident benefits on resource and en-
ergy usage within the network. If sensors measure spatially correlated
values, values collected from a subset of nodes can be used to materialize
the uncollected values from other nodes [20, 21]. Similarly, temporally-
correlated values may be collected infrequently and missing values can
be interpolated [15, 22]. By placing models in the mote itself, the mote
may transmit model parameters in lieu of the data, compressing or sup-
pressing entirely the data stream [23-25].
Thereisalsoworkondefiningacommonconceptualmodelofevent
processing based on event driven architectures [27] and event processing
networks [28]. In PCA model [11], the notion of event history or event
flow is different from those used in [19, 20] such that the event history
flow takes embedded uncertainty. In fact it contains observations (event
clusters), which consist of multiple possible events. In those models
an event history itself is considered deterministic and the uncertainty
on event history is expressed as there can be multiple possible event
histories. Due to this difference, the rule semantics is totally different
from the conditional representation in [19]. Ganeriwal et al. [26] dis-
cuss the reputation-based framework for high integrity sensor networks.
The model evaluates the trustworthiness of the nodes and various mis-
behavior types of nodes in the network. The model uses the Bayesian
formulation and updates the trust with direct and indirect trust calcu-
lations.
2.2 Sensor Event Detection
Much work has been done in sensor networks on composite event
detection. Directed Diffusion [29] is among the earliest event-based ap-
proaches. In this approach, a node would request data by sending inter-
ests, which is conceptually similar to subscriptions in a publish/subscribe
system. Data found to match those interests are then sent towards that
node. A different framework based on event classification is the Online
State Tracking [30] approach. This technique consists of two phases:
the first phase is the learning process where new sensor readings are
classified to states, and the second phase is the online status monitor-
ing phase where nodes are collaborating to update the overall status of
the network. The work is quite unique in the sense that it moves away
from individual node readings and views the whole network as a state
machine.
