Citation

Okayama Y. Fire Safety J. 1991; 17(6): 535-553.

Copyright

(Copyright © 1991, Elsevier Publishing)

DOI

unavailable

PMID

unavailable

Abstract

Various fire detection methods have been proposed worldwide, but it is quite difficult for them to discriminate between real fire in its incipient stage and nongenuine fire phenomena which often occur.Lately, extensive studies on neural nets which imitate neural circuits of the human brain and work in computers like brains have been made in many fields. These neural nets are capable of handling pattern-matching of analogue input and output which has so far been said to be difficult. Furthermore, more precise fire discrimination can be expected if they are adapted to the process of fire detection judgement.For evaluation of performance of the neural net formed with three layers, i.e. the input, hidden and output layers, analogue input data and results were given to the net which in turn learnt the relation by a back-propagation method. The definitions of input and output were stored as strength of strings between two layers of the net. Thus, the net can obtain results very close to expectation from what has been learnt even with an undefined combination of input values. Unlike a conventional pattern-matching method, it is capable of giving answers to any inputs without fail. In addition, the learned neural net does not give such answers as would differ so much from the practical use, even if there aren't a sufficient number of patterns in the definition table. It also requires less memory size and time to make decisions. Therefore, it will be easy to put this neural net into practice in the future.