High-Dimensional Time Series Feature Extraction for Low-Cost Machine Olfaction
Pratistha Shakya, Eamonn Kennedy, Christopher Rose, Jacob K. Rosenstein
Abstract
The complexity of airborne odors offers interesting challenges and opportunities for chemical detection and identification. Biological olfactory systems have evolved to extract information from spatiotemporally complex odor plumes, but many engineered electronic noses use only coarse time features while neglecting valuable transient fluctuations. In this paper, we use the TruffleBot, our low-cost e-nose platform, to dynamically `sniff' odors while collecting multidimensional chemical, pressure and temperature time series. By extracting high-dimensional time series features (TSF) from a diverse set of relatively low-bandwidth sensor signals, we can identify subtle differences in odor concentration and composition. We use this approach to perform a variety of classification experiments, including the discrimination of three similar beers at >98% accuracy. Additionally, we demonstrate that time series features can be aggregated and applied to improve concentration estimation of ethanol by a factor of three, to the limit of our experimental calibration error.