Back-propagation-based neural network with a two sensor system for monitoring carbon dioxide and relative humidity

The delicate problem of independently determining the concentration of carbon dioxide and relative humidity in a gas mixture is solved by recording the signals of two quartz microbalances coated with a functionalized sensitive polymer layer. The data were analysed by back-propagation-based neural networks. We tested two different architectures, a one-stage net and a two-stage net (which consist of two network structures in series) with respect to their generalisation ability. The one-stage network calculated the concentration of carbon dioxide with an error less than 12% and that of relative humidity with an error less than 5%. The two-stage network recognised in the first step seven categories of relative humidity with a success rate of 100%, while the maximal error of carbon dioxide concentration calculated in the second level of this architecture is reduced to 9%. In addition, the generalisation ability of the two-stage network is improved compared to the one-stage network.