Odor sensing system with multi-dimensional data analysis

Abstract

There are two current approaches in odor sensing systems. One is an odor biosensor using actual olfactory receptors. Although there have been many artificial odor sensors proposed over the past three decades, there is still room for improvements in terms of sensitivity, selectivity and stability. A biomimetic approach using olfactory receptors of a living body is expected to enhance capability. The second is a deep learning technique to predict odor impressions. The mapping of mass spectra onto sensory test data based on the semantic differential method was performed. A method of using two autoencoders for both independent- and dependent-variable spaces together with ordinal multi-layer perceptrons was proposed. Its classification accuracy was better than that of the conventional regression method. Moreover, several cost functions were applied to the autoencoder and evaluated. It was found that the Itakura-Saito cost function was superior to others for reproducing small peaks in the mass spectrum. The author believes these are key technologies for realizing a sophisticated olfactory sensor.