Toward Accurate Odor Identification and Effective Feature Learning With an AI-Empowered Electronic Nose

The development of Internet of Thing technology and robotics can be significantly promoted to a higher level with a precise digital sense of odors. However, the detection and identification of diverse odors using electronic sensor systems pose significant challenges. Electronic nose (E-Nose) based on gas sensors provides a cost-effective solution to detect odors. While previous research has primarily focused on enhancing the discriminative capability of E-Noses through machine learning techniques, limited attention has been given to the quality of features extracted or learned from E-Nose data. This article presents a comprehensive study of E-nose design, digital odor measurement, and feature analysis methods in classifying many different odors into distinctive digital signatures. Experimental investigations involved the construction of a novel E-Nose system with automated data processing capabilities, enabling the study and discrimination of various odors, including essential oils, coffee, and whiskey. In the realm of data analysis, multiple feature extraction methods were compared and evaluated, including 1D-convolutional autoencoder (C-AE) and 1-D convolutional neural network (1D-CNN). Motivated by recent progress in representation learning within the realm of face recognition, particularly its proficiency in generating distinctive features within the angular space, an algorithm incorporating a 1D-CNN model complemented by ArcLoss was developed. This innovative approach resulted in an exceptional classification accuracy of 97.76%, demonstrating its robustness in identifying and distinguishing a complex array of odors originating from a variety of essential oils.