Time Series Feature Extraction for Machine Olfaction

Biological olfactory systems routinely process complex transient odors and sensory inputs, while today's electronic noses often leverage only coarse time features. In this work, we apply generalized time series feature extraction to improve the performance of the TruffleBot platform. TruffleBot is a low cost electronic nose which `sniffs' to collect multidimensional chemical, pressure and temperature time series. By learning from a larger space of time series features, such as autocorrelations, we enable TruffleBot to identify odors in spite of differences in transient concentration fluctuations, and to classify the odors of three similar beers at >98% accuracy (74 out of 75 trials). Additionally, we show that learned time series features can enable ethanol vapor concentration estimation within ±2.3%, near the limit of our experimental error.