Application of unsupervised clustering methods to the assessment of malodour in agriculture using an array of conducting polymer odour sensors

Odour sensing instrumentation based on arrays of sensors that display broad specificity and high sensitivity to odorous chemicals have been developed. Methods of unsupervised reduction of complex multidimensional data were examined, and useful algorithms were adapted for the use with broad specificity sensor arrays. As a result, a simple interface that would allow the human observer to decide easily whether a particular odour pattern could be distinguished from another on the basis of Euclidean distances between patterns, and calculation of 95% confidence limits around individual clusters of data was developed. The best performance was obtained from a combination of principal components analysis used as a starting point for Sammon mapping. This combined the invariance of the eigenvector calculation with the Euclidean distance mapping of the Sammon procedure, without the disadvantages of rotation of clusters when the order of patterns in a database was changed. The methods were applied to the assessment of odour differences of fresh pig slurry from pigs fed with different diets. The results show that the transformation of multidimensional data into two-dimensional clusters allowed easy visualisation of the difference in odour between slurries from the pigs fed with different diets.