A sensor conditioning principle for odor identification

This work formulates a sensor array optimization scheme for odor identification. It hinges on a performance index widely used in the signal theory, namely the Mahalanobis distance, which gives a solid quantification of the separability among odor classes. Optimizing this index over the controllable operating parameters of the sensor array minimizes the response variability within each class and simultaneously maximizes the spread of the class prototypes (i.e., the class centers) in the feature space. The evaluation of the empirical measure is data-driven, yet simple enough to be performed on-the-fly, provided that a representative set of labelled measurements from each class is available. We finally demonstrate on a sample dataset that tuning the temperature of a metal-oxide sensor array based on the suggested criterion yields a substantial improvement in the classification performance.