Aplicação de análises estatística e neural para reconhecimento de sinais de odores

Abstract

This work investigates the use of an electronic nose prototype for prognosis of Diabetes Mellitus. The work involves five main parts: (1) building of odors database by aroma sensors; (2) an evaluation of the odors database through multivariate statistics techniques;(3) use of Multilayer Perceptron (MLP) and Radial Basis Function (RBF) Artificial Neural Networks (ANN) in the odors recognition; (4) an evaluation of statistics and connectionist approaches; and (5) a case study. The multivariate statistical method used was the Principal Components Analysis (PCA) as characteristic extraction, beyond of an evaluation of performance for Hypothesis Test. Also had been used the K-Nearest-Neighbour (KNN) classifier for classification of odors signs. The connectionist approach involves the MLP e RBF ANN trained with de odors database. For the ANN’s inputs were used the Normalization as pre-processing of data and PCA for characteristic extraction and reduction of data dimensionality. The signals generated by the four sensors of an electronic nose prototype exposed to the urine from the diabetics and non-diabetics patients and the characteristics signals of the sensors without the presence of volatiles substances were analyzed. The use of characteristic extraction by PCA improved the performance of the classifiers used. In the experiments, theKNN classified correctly all the data set from the database divided in training and test after PCA. The better MLP ANN obtained a classification of 77% to the norma lized data without PCA use and of 100% after PCA while the RBF ANN obtained a classification of 66% to the normalized data and the 100% after PCA in the odors classification.