Database Reference
In-Depth Information
A lighter passive version of this approach, called PAD [65], is described
in more recent literature. PAD uses a variant of belief networks to infer
the most likely cause of communication misbehavior symptoms. It is
a passive approach in that the diagnostic system does not introduce
communication trac of its own, but rather piggybacks status bytes on
existing trac. It is light in that the system restricts network status
monitoring overhead to only two bytes per packet. These bytes, called
a
mark
, simply carry the ID of some intermediate node on the packet's
path to the collection base-station, as well as the hop count from the
intermediate node to the source. By cleverly decising which node should
mark which packet, network topology information can be reconstructed
at the base-station, and diagnostics can be performed to identify the
likely causes when anomalous changes occur.
2.2 Supervised Classifiers
Two different examples of supervised learning approaches in sensor
network debugging literature are the Diagnostic PowerTracer [50] and
the SNTS debugging tool [51]. The former learns signatures of different
known failures
reproduced intentionally during laboratory testing, such
that it could later recognize these failures in the field. The latter learns
the distinguishing features of newly encountered
unknown failures
by
contrasting them with normal behavior. These features then provide
clues regarding the root cause.
The Diagnostic PowerTracer was motivated by the need to remotely
identify root causes of failures of
silent nodes
. While techniques that
localize the failed node [78, 65] can pinpoint which node died, they of-
fer less insight into the reason for its silence after the node stops re-
sponding. To circumvent this challenge, the PowerTracer features an
external hardware module that samples (at the rate of a few Hz) the
power consumed by each sensor node, and sends those samples via a
low-power low-bandwidth radio to a diagnostic base-station. The base-
station quantizes the reported power values in the trace into discrete
symbols (called power consumption
states
). It then derives from the
sequence of symbols in the trace a probabilistic state transition diagram
and matches it against those of known failures to identify the problem
with the silent node. The approach was shown to correctly differenti-
ate between OS crashes, radio failures, antenna failures, water-induced
electrical failures, and battery depletion, among other failure states.
A different supervised classification-based approach is exemplified in
the SNTS debugging tool [51]. The tool uses the PART [28] algorithm
from the Weka data mining library [70, 36] as the underlying classifer to
