Self-verifying time-resolved multicomponent gas analysis from raw sensor signals

Abstract

Gas sensor arrays can estimate the composition of gas mixtures from their response patterns. However, when a sensor is used outside the laboratory, the true gas concentrations are unknown, so there is usually no direct way to determine whether the estimated result is reliable. Here, we develop a physics-constrained neural network that estimates time-varying gas concentrations and reconstructs the corresponding sensor signals at the same time. We tested the method with a twelve-channel membrane-type surface stress sensor array exposed to water-ethanol mixtures whose concentrations changed over time. The model reconstructed the concentration profiles of both components from the raw signals. Artificial offsets and sensitivity changes in one channel were detected from disagreement between the measured and reconstructed signals. Deletion or replacement of part of a measurement record was detected from violations of the sorption and mechanical equations. This approach enables time-resolved multicomponent gas analysis with self-verification from a single measurement.