Neural processing in an electronic odour sensing system

The past ten years has seen a significant increase in activity in the application of multisensor arrays to odour classification and description. Much of this work has been based around systems consisting of a semiconductor gas sensor array, elementary signal conditioning and microcomputer based pattern recognition (P. Corcoran, 1993). Most interest in the area of pattern recognition has been in the use of multivariate statistical methods. These techniques were applied to the discrimination and classification of a number of odours including alcohols and beers by H.V. Shurmer et al. (1990) and coffees by J.W. Gardner et al. (1992). Significant success has been achieved in these studies, which were primarily based upon the steady state response of multisensor arrays in controlled environments. However, to translate this success to real applications in noisy environments requires a new set of processing tools to accommodate sensor drift, nonlinearity, and variations in the steady state response that arise when the multisensor array is present in atmospheres containing interfering measurands. A feasibility study is described that uses neural networks and multivariate statistics for the discrimination of wine aromas. Discriminant function analysis is compared to the use of a 3 layer MLP architecture when applied to the classification of five wines. Self organising maps and their scatter plots, derived using a Kohonen network are presented. This analysis shows the distribution of the wine aroma data. The relationships between this technique and the supervised techniques, as applied to sensor systems, are also discussed