A smooth neural approximator for multi-criteria fire detection

CASAGRANDE DANIEL; KAYHAN FARBOD; SEMESHENKO VIKTORIYA; ZAMBON CRISTIANO

University of Duisburg, Germany

Congreso; 13th International Conference on Automatic Fire Detection (AUBE '04); 2004

Multi-criteria fire detectors base their decisions, of whether to alarm or not, on signals read by different types of transducers. Many different algorithms have been designed for merging all of the input data in order to obtain a decision as close as possible to the intended one. However, these kinds of algorithms can only approximate the real alarm region in a segmented way, i.e. dividing the original region in many parts and using some tricks to interpolate the surface. A different approach in algorithm design is the neural approximation. Unfortunately, this kind of approach has an intrinsic over-training possibility that formally reduces the practical realization. In reality, there is no way to guarantee that the error surface has no peak outside the set of points used for the validation test. On the other hand, there is a widely known particular kind of neural network, the perceptrons that does not require a learning process. The aim of the research presented herein is to understand if it is possible, under some conditions, to use perceptrons to define the alarm region. If this could be done, the stability of the detector would be assured since each perceptron is intrinsically stable. We conclude that if the alarm region is convex than it can be approximated with a given accuracy provided that a sufficient number of neurons are used. Otherwise, if the alarm region is not convex, an algorithm has been designed to divide it into convex volumes and to merge the outputs of all perceptrons.