- Author
-
Shaffer, R. E.
|
Rose-Pehrsson, S. L.
|
Williams, F. W.
|
Barry, C.
|
Gottuk, D. T.
- Title
- Development of an Early Warning Multi-Criteria Fire Detection System: Analysis of Transient Fire Signatures Using a Probabilistic Neural Network.
- Coporate
- Naval Research Laboratory, Washington, DC
Hughes Associates, Inc., Baltimore, MD
- Report
-
NRL/MR/6110-00-8429,
February 16, 2000,
32 p.
- Distribution
- AVAILABLE FROM National Technical Information Service (NTIS), Technology Administration, U.S. Department of Commerce, Springfield, VA 22161. Telephone: 1-800-553-6847 or 703-605-6000; Fax: 703-605-6900. Website: http://www.ntis.gov
- Keywords
-
fire detection systems
|
warning systems
|
neural networks
|
fire detection
|
sensors
|
smoke detectors
|
pattern recognition
- Identifiers
- sensor data fusion; sensor array; multivariate analysis
- Abstract
- Early detection of fires is a critical component of the Office of Naval Research's Damage Control: Automation for Reduction Manning (DC-ARM) program. The research described in this report is a continuation of a collaboration between the Naval Research Laboratory (NRL) and Hughes Associates to develop and evaluate an early warning, multi-criteria fire detection system for use in DC-ARM. In this program, Hughes Associates collected a large database containing the signatures from real and nuisance alarm sources for several different types of sensors. Our approach to early fire detection is based on the premise that a combination of sensor technologies coupled with pattern recognition methods could provide for faster, more accurate fire detection than any single sensor technology that measures a physical quantity (e.g., particles or heat) or a fire decomposition vapor (e.g., CO2, 02). Previous research in the analysis of these fire signatures produced several combinations of sensors that could provide increased detection sensitivity, decreased detection time and improved nuisance source false alarm rejection. Those experiments assumed that a baseline sensor reading could be obtained for the smoke detectors. The work described in this report takes a slightly different approach. Here we assume that it will not be possible to determine an appropriate baseline level and all analyses are performed on the raw sensor outputs. This makes the fire detection problem more difficult since the day-to-day variation in sensor readings is not removed Using this raw data, we investigated several data analysis issues critical to developing an early warning, multi-criteria fire detection system including: 1. the importance of the sensor rate of change (i.e., slope); 2. the best sensor combinations; 3. earliest possible tire detection times; and 4. the optimal procedure for training the probabilistic neural network. In this study, the large fire signature database described in reference 2 was used. This database consists of data from twenty different sensors for 88 fire events and 38 nuisance sources. Table 1 provides a list of the twenty sensors used in this study. Because the MIC sensor produces three different sensor readings, a total of 22 sensor outputs were studied. During preliminary investigations it was discovered that the ODM sensor stopped working during the middle of the four experiments (DCAS012, DCASOl5, DCAS018, and DCASO19). During DCASO54 and DCAS057, the photoelectric smoke sensor was not operation.