Neural Networks for Smart Fire Detection. Final Report.
Neural Networks for Smart Fire Detection. Final Report.
(7051 K)
Milke, J. A.; McAvoy, T. J.
NIST GCR 96-699; 317 p. December 1996.
Sponsor:
National Institute of Standards and Technology,
Gaithersburg, MD
Available from:
National Technical Information Service
Order number: PB97-138267
Keywords:
fire detection; experiments; data analysis; light
obscuration
Abstract:
Research was conducted using multiple sensors with an
algorithm to detect fires more quickly than currently
available smoke detectors while also decreasing the
susceptibility to unnecessary alarms. The effort
involved the production of signatures from three types
of sources: flaming fires, non-flaming fires and
non-fire, nuisance sources, followed by analysis to
recognize signature patterns for the three types of
sources. The first phase of research consisted of
establishing the feasibility of distinguishing between
signatures from fire and non-fire sources using a
small-scale apparatus. The second phase consisted of
introducing the signatures in a 12 ft. square room with
a height of 8 ft. Measurements included CO, CO2, and O2
concentrations, presence of oxidizable gases, light
obscuration and temperature. The signatures measured
could be associated with the three types of sources.
Using a multivariate statistical analysis, the response
time of a prototype detector was appreciably less than
that of commercially available detectors, with a
significant reduction in unnecessary alarm
susceptibility. In the third phase, pairs of sources
were provided simultaneously to determine if a nuisance
source could mask the signature from a fire source and
if two nuisance sources provide a signature similar to
that from a fire. Results indicate that the ratio of
the CO to CO2 concentrations is representative of
flaming fire sources and to a limited extent for
non-flaming fire sources, independent of the presence of
a nuisance source.
