Comparação estatística de performance de métodos de redes neurais para sistema de olfação biológica

Abstract

One of the human senses that has several aspects get to be elucidated is the olfactory sense. Therefore, many scientists have been studying this sense in order to better understand how does the information processing happens until the brain recognize it. There were lots of theories regarding olfactory system functioning, in which its authors try to explain how the reception, the analysis and the odor detection occur. Many people still use their own noses as a working tool. In this kind of job, people are trained to inhale and detect different odors. It is considered as an exhausting and risky job for those professional that, for example, could inhale toxics gases. In order to solve this problem, many systems that try to simulate a biological nose were developed. These systems are known as artificial noses. An artificial nose is an equipment composed of sensors and a pattern recognition system. The sensors are responsible for detecting odor signs from the external environment. The pattern recognition system is used to classify the signs sent by sensors and to provide a result from these signs.In the present work, artificial neural network techniques were used for the patternrecognition process, once these techniques are non-parametric and usually nonlinear. The usage of artificial neural networks as an odor recognition system has been quite advantageous. These networks are capable of working with non-linear data and also have an adaptation capability, they are tolerant to errors and noise, and have parallel processing. MLP, RBF e PNN were used in the development of an odor recognizing system based on a biologic system model and its results were compared, using the Wilcoxon test on the respective network models without the adaptation to the biologic model.