Active wavelength selection for mixture analysis with tunable infrared detectors

This article presents an active wavelength selection algorithm for multicomponent analysis with tunable infrared sensors. Traditional techniques for wavelength selection operate off-line; as a result, the resulting feature subset is fixed and only optimal for the specific mixtures and noise levels in the training set. To address this limitation, the proposed algorithm interleaves the wavelength-selection and sensing steps so that the feature subset adapts to information from previous measurements. At each point in the process, the algorithm maintains a pool of candidate solutions (i.e., mixtures) consistent with all past measurements, then selects the wavelength that maximizes discrimination across the pool. The algorithm uses a weighting function based on Akaike information criterion to promote parsimonious solutions and balance exploration vs. exploitation strategies. The algorithm is validated experimentally on binary mixture problems with a tunable infrared detector (Fabry-Perot interferometer), and its performance on higher-order mixtures characterized in simulation with a large spectral library. Active wavelength selection outperforms passive strategies, particularly at low signal-to-noise and foreground-background ratios, and when mixture components are similar, in which case the problem becomes ill conditioned. (C) 2014 Elsevier B.V. All rights reserved.