Bayesian approaches to assigning the source of an odour detected by an electronic nose.

After a brief review of electronic nose technology, the use of an Australian electronic nose to identify an unknown odour out of a set of known odours is described. Multivariate supervised learning is accomplished by applying Bayes' theorem to data from metal oxide semiconductor sensors responding to each of a number of target odours. An odour from an unknown source is then assigned a probability of membership of each of the training sets by applying either a Naïve Bayes algorithm to the deemed independent data from each sensor, or to a multinormal distribution of the data. A flat prior (equal probabilities of each outcome) is usually adopted, but for particular situations where one odour is known to predominate, then suitably weighted priors can be used. A source 'none of the above', which has a small likelihood covering the space of the possible sensor responses, is included for completeness. This also avoids the assignment to a source that has an extremely small probability but which is greater than that of any other source. Examples are given of a single source (detecting diabetes from a patient's breath), and three sources of unpleasant odours in a meat processing plant. The 18th century cleric Thomas Bayes gave his name to an elegant statement of the probability of an event – in this case identification of an odour – given some evidence: output from a number of metal oxide semiconductor sensors. Knowing the distributions of outputs for target odours, we assign the probabilities of an unknown odour. The greatest probability wins! (Image credit: E-nose Pty Ltd and D. B. Hibbert.) This article belongs to the 10th Anniversary Collection of RACI and AAS Award papers. [ABSTRACT FROM AUTHOR]