<title>Application of cascade correlation learning network and tin oxide sensor arrays to the analysis of multicomponent gas mixtures</title>

Abstract

ABSTRACT This paper describes an approach for determining the individual analyte concentrations in a mixture ofgases using array tin oxide sensors and cascade correlation learning network. In the present approach asmall array of six semiconductor oxide (Sn02:Pd, Pt) sensors is used to recognize mixtures of air withfour pollutants: carbon oxide, methane, propane/butane and methanol vapours. The cascade-correlation algorithm offers a reasonably small net, which is built automatically. Keywords: Neural networks, Gas sensors array, Combustible gases.1. INTRODUCTIONThere are many instanses where it is necessary or desirable to determine concentrations of individualanalytes in a mixture of gases. For example, this information may be necessary to control a chemicalprocess or to monitor the safety ofgases in an underground mine. The traditional approach to this sensingproblem has been to built individual sensors that respond exclusively to each individual analyte in themixture. Hence, each sensors output corresponds to the concentration of a particular analyte. Catalyticdevices and gas sensors (semiconductor oxides) are widely used as convenient means of estimating offlammable gases in air. One of the main problem with this type of gas sensors is their lack of selectivity orpartial selectivity. Although these sensors are reputed not to be selective, a certain degree of selectivitycan be introduced either by operating at different temperatures, or by using catalysts. Another attractiveway to achieve this selectivity is the use of an array of sensing elements and, thanks to pattern recognitiontechniques [1], the identification of an unknown gas or vapour. In sensor applications, artificial neuralnetworks have already been applied to recognize of an multi component mixtures of gases [2},[3],[4},[5}.In this paper, in order to recognize four combustible or toxic gases pollutants (carbon oxide, methane,propane/butane and methanol vapour), a small array of six semiconductor oxide (Sn02 :Pd, Pt) sensorswith various compositions is used. This array of the sensors was exposed to various mixtures of air withthese four pollutants. The approach presented in this paper uses the artificial neural network type CCLA(Cascade-Correlation Learning Architecture) to learn the relationships between the sensor outputs andinput analyte cocentrations.