Exploring the role of feature subset combinations on performance of multisensor fire detection

Abstract

Multisensor technology has been widely accepted as the next generation fire detection technology. The objective of this research is to explore influence of various feature subset combinations on multisensor fire detection performances. ANN models of BP, RBF, and PNN are selected as fire detection classifier. Four fire signatures, namely temperature, smoke obscuration, CO, and CO2 concentrations, are used to generate possible fire detection combinations. Fire detection performance of reliability and sensitivity are investigated of different combinations using parameters of wrong alarm rate and detected fire point respectively. RESULTS indicate that fire detection performance declines with increasing number of ANN in multisensor fire detection. Various combinations with the same ANN number can still produce dramatically different multisensor fire detection performance, and ANN models can overcome the disadvantage of various combinations to a large extent. Performance of PNN is highlighted of all the three selected ANN models by investigation results.