A feature selection strategy for the analysis of spectra from a photoacoustic sensing system

In the frame of the EU project CUSTOM, a new sensor system for the detection of drug precursors in gaseous samples is being developed, which also includes an External Cavity-Quantum Cascade Laser Photo Acoustic Sensor (ECQCLPAS). In order to define the characteristics of the laser source, the optimal wavenumbers within the most effective 200 cm -1 range in the mid-infrared region must be identified, in order to lead to optimal detection of the drug precursor molecules in presence of interfering species and of variable composition of the surrounding atmosphere. To this aim, based on simulations made with FT-IR spectra taken from literature, a complex multivariate analysis strategy has been developed to select the optimal wavenumbers. Firstly, the synergistic use of Experimental Design and of Signal Processing techniques led to a dataset of 5000 simulated spectra of mixtures of 33 different gases (including the 4 target molecules). After a preselection, devoted to disregard noisy regions due to small interfering molecules, the simulated mixtures were then used to select the optimal wavenumber range, by maximizing the classification efficiency, as estimated by Partial Least Squares – Discriminant Analysis. A moving window 200 cm -1 wide was used for this purpose. Finally, the optimal wavenumber values were identified within the selected range, using a feature selection approach based on Genetic Algorithms and on resampling. The work made will be relatively easily turned to the spectra actually recorded with the newly developed EC-QCLPAS instrument. Furthermore, the proposed approach allows progressive adaptation of the spectral dataset to real situations, even accounting for specific, different environments.