Quantification of Gas Mixtures with Active Recursive Estimation

We present an active-sensing strategy to estimate the concentrations in a gas mixture using temperature modulation of metal-oxide (MOX) sensors. The approach is based on recursive Bayesian estimation and uses an information-theoretic criterion to select operating temperatures on-the-fly. Recursive estimation has been widely used in mobile robotics, e.g., for localization purposes. Here, we employ a similar approach to estimate the concentrations of the constituents in a gas mixture. In this formulation, we represent a concentration profile as a discrete state and maintain a 'belief' distribution that represents the probability of each state. We employ a Bayes filter to update the belief distribution whenever new sensor measurements arrive, and a mutual-information criterion to select the next operating temperature. This allows us to optimize the temperature program in real time, as the sensor interacts with its environment. We validate our approach on a simulated dataset generated from temperature modulated responses of a MOX sensor exposed to a mixture of three analytes. The results presented here provide a preliminary proof of concept for an agile approach to quantifying gas mixtures.