Quantification of gas concentrations in mixtures of known gases using an array of different tin-oxide sensors

Abstract

Two multilayer-perceptron (MLP) artificial-neural-network (ANN) committee methods and a mathematical-macromodelling method for determining the concentrations of gases in a known gas mixture from the outputs of an array of tin-oxide gas sensors placed in the mixtures are described. The committee approach is also used to determine the associated error bars. A large set of artificial training data generated from the small set of experimental data was used to train the MLPs. For the Bayesian-trained committee methods average predicted concentration errors of 1.66 to 9.49% were obtained. The macromodelling method resulted in errors of 19.7 to 33%, but was much easier to implement and faster. Training by back propagation gave much worse accuracy. The average calculated error bars were in good agreement with the actual errors in prediction. The concentration errors were comparable to those yielded by other methods and were at least partly determined by the original errors in the concentration measurements.