DeepSniffer: A meta-learning-based chemiresistive odor sensor for recognition and classification of aroma oils

Abstract

A meta-learning algorithm, conventionally used for visual recognition, was applied to the recognition and classification of aroma oils. A printable chemiresistive sensor array was fabricated, based on composites of carbon black with various active materials. Standard aromatherapy kits with 30 types of essential oils were used as targets in an odor sensing experiment. Benefiting from the pattern recognition ability of the fabricated sensor array, a high-quality dataset was obtained with 30 aroma oil classes, in which each class had nine replicate samples. A deep metric learning model, based on a Siamese neural network and a multilayer perceptron, was used to perform the N-way k-shot meta-learning. A test accuracy of over 98.7% was obtained for 31-way 9-shot learning, on discriminating whether the input pair samples were taken from similar or dissimilar classes. The model was effective in extracting meta-features of the aroma oils; this was proved by the improved clustering effect of samples in the spaces of principal components analysis and t-distributed stochastic neighbor embedding. The 30 aroma oils were divided into two datasets according to 6-fold cross-validation: 25 aroma oil classes (plus one blank class) as seen classes for constructing 26-way 9-shot learning models and the remaining five aroma oils as unseen classes for prediction. Average accuracies of 93.5% and 93.9% were achieved for recognition of the unseen aroma oils from the seen classes and classification of the unseen aroma oils themselves, respectively, demonstrating the effectiveness of the developed sensor and model for odor recognition and classification.