Data-centric artificial olfactory system based on the eigengraph

Abstract Recently reported studies of electronic nose systems that combine chemiresistive gas sensors with computational analysis technologies tend to waste significant amount of important data. Since the sensitivity-oriented data composition have made it difficult to discover meaningful and valuable data to apply artificial intelligence analysis in terms of in-depth analysis of odor attributes specifying the gas identities, ultimately resulting in hindering the development of artificial olfactory technology. Here, we propose a novel data-centric approach for realizing a standardized artificial olfactory system inspired by the human olfactory mechanism. We designed an olfactory receptor-like sensor array in which each channel has independent gas sensing characteristics through semiconductor processes. In addition, the physicochemically optimized sensor array reproducibly generated eigengraphs with highly refined waveforms which contribute keeping invariability of input space to deep learning architecture. The implicit odor attributes of the time domain eigengraphs were mathematically represented by the newly defined frequency domain feature vectors. The effectiveness of the MFCC feature vectors in deep learning for gas classification was clearly demonstrated in terms of training time reduction and inference performance improvement despite the extremely decreased amount of training data. We suggest that our strategy can be widely applied as a source technology to develop standard artificial olfactory systems.