Tuesday, October 31, 2006

Paper: "Fidelity and Yield in a Volcano Monitoring Sensor Network"

Fidelity and Yield in a Volcano Monitoring Sensor Network
Geoff Werner-Allen and Konrad Lorincz, Harvard University; Jeff Johnson, University of New Hampshire; Jonathan Lees, University of North Carolina; Matt Welsh, Harvard University


We present a science-centric evaluation of a 19-day sensor network deployment at Reventador, an active volcano in Ecuador. Each of the 16 sensors continuously sampled seismic and acoustic data at 100 Hz. Nodes used an event-detection algorithm to trigger on interesting volcanic activity and initiate reliable data transfer to the base station. During the deployment, the network recorded 229 earthquakes, eruptions, and other seismoacoustic events.

The science requirements of reliable data collection, accurate event detection, and high timing precision drive sensor networks in new directions for geophysical monitoring. The main contribution of this paper is an evaluation of the sensor network as a scientific instrument, holding it to the standards of existing instrumentation in terms of data fidelity (the quality and accuracy of the recorded signals) and yield (the quantity of the captured data). We describe an approach to time rectification of the acquired signals that can recover accurate timing despite failures of the underlying time synchronization protocol. In addition, we perform a detailed study of the sensor network's data using a direct comparison to a standalone data logger, as well as an investigation of seismic and acoustic wave arrival times across the network.


Blogger Shanth said...


Geoff Werner-Allen presented this
science-centric evaluation of a 19-day
sensor network deployment at Reventador,
an active volcano in Ecuador. The data
collected by the sensor network
deployment was evaluated based on five
metrics, namely, robustness, event
detection accuracy, data transfer
performance, timing accuracy and data
fidelity. Of these, Geoff dealt with
robustness, timing accuracy and data
fidelity in his talk.

The sensor hardware they used for
Volcano monitoring was small and
provided real-time data acquisition,
unlike the conventional standalone
dataloggers that were unwieldy and
logged data to a local flash drive.
Their sensor network contained 16 sensor
nodes, each equipped with seismometer,
microphone and antenna. These nodes
continously sample seismic and acoustic
signals and log the data to a local
flash memory. They also run an event
detection algorithm that transmits a
time-stamped report to the base station
upon detection of a seismic event. The
base station, located 4.6 km away from
the sensor deployment, initiates data
collection if it receives triggers from
more than 30% of the sensor nodes within
a time interval.

The overall robustness of the system was
limited by power outages at the base
station and a single three-day software
failure. Discounting these, the mean
node uptime exceeded 96%, indicating
that the sensor nodes themselves were
reliable. Flooding Time Synchronization
Protocol (FTSP) was used for
time-synchronization between sensor
nodes. Although pre-deployment results
with FTSP was good, during deployment,
they ran into stability issues leading
to occasional incorrect time-stamp
reports. They developed a time
rectification approach that filters and
remaps recorded timestamps to accurately
recover timing despite the incorrect
timestamps. They evaluated the fidelity
of the collected data by performing an
analysis of the seismic and acoustic
signals from a seismological
perspective. Their results indicate that
the collected signal quality and timing
matches the expectation of the
infrasonic and seismic activity produced
by the volcano.

Geoff concluded his talk with the three
lessons that they learnt from this
deployment: ground truth and
self-validation are critical, network
infrastructure is more brittle than
sensor nodes and its important to build
confidence with domain scientists.

During the Q&A session, there was a
question on whether they look at sensor
networks as a tool that scientists in
other domains could use without
requiring the computer scientists'
presence. Geoff responded by stating
that it was a great observation and that
they wanted sensor networks to
eventually be used like that. Someone
from SUNY-Stony Brook asked if it was
possible to simply broadcast time from
the base-station. Geoff replied that
such a broadcast would work only with
single hop networks, which isn't the
case with the Reventador deployment.
Mehul Shah (HP Labs) asked about the
fidelity of the measurements with
respect to its relevance to the domain
scientists. Geoff said that the
scientists are still working on the
results obtained and that their initial
observations are encouraging.

1:30 PM  

Post a Comment

<< Home