The development of algorithms for a smoke detector with neuro-fuzzy logic

• Published in: Fuzzy sets and systems 1996, Vol.77 (2), p.117-124
• Main Author: Thuillard, Marc
• Format: Article
• Language: English
• Subjects: Data analysis; Fuzzy logic; Neural networks; Sensors
• ISSN: 0165-0114; 1872-6801
• DOI: 10.1016/0165-0114(95)00050-X

We have developed a neuro-fuzzy competitive learning procedure that permits the automatic evaluation of the number and form of the fuzzy rules and membership functions necessary to describe a set of data. The input space of dimension k is divided into 2 k hypercubes with a modified Kohonen's algorithm. The architecture of the learning algorithms is hierarchical. New membership functions and fuzzy inference rules are introduced according to the results of the classification on the previous level. After learning, the inference rules are simplified using boolean logic. This method reduces the required computing time by limiting learning to the domains in the input space data that could not be associated to a unique class of events. We have applied this method to the development of algorithms in a linear smoke detector. The signal from the light receiver of the smoke detector is given to a battery of digital filters, each filter extracting one characteristic feature from the signal. The filtered signals are then used as input data for the fuzzy system. The algorithms, we developed, are able to distinguish between smoke and deceptive phenomena in order to suppress unwanted alarms.