Analysis of neural networks and analysis of feature selection with genetic algorithm to discriminate among pollutant gas

A multisensor based on tin and tin–titanium oxides has been utilised to detect pollutant gases (NO2, CO, toluene and octane). The sensitive layers are deposited by r.f. reactive sputtering. Some tin oxide sensors are doped with Pt. Measurements are carried out with single gases and gas mixtures (two and three gases) in dry air at 250 ◦C. An exhaustive analysis of several networks and feature extraction and selection is done to discriminate among four different pollutant gases. First the sensor responses are analysed with principal component analysis (PCA). The results are not good enough for mixtures. Then several pre-processing techniques and several artificial neural networks (ANN) are studied. Two models of neuronal networks are used: probabilistic neural network (PNN) and multilayer perceptrons (MLP). A selection of the sensors and of pre-processing techniques was made with a genetic algorithm (GA). © 2004 Elsevier B.V. All rights reserved.