Partial least squares regression as a tool to retrieve gas concentrations in mixtures detected by using quartz-enhanced photoacoustic spectroscopy

We report on a statistical tool based on partial least squares regression (PLSR) able to retrieve single-component concentrations in a multiple-gas mixture characterized by absorption features spectrally overlapping. Absorption spectra of mixtures of CO-N2O and mixtures of C2H2-CH4-N2O, both diluted in N2, were detected in the mid-IR range by exploiting quartz-enhanced photoacoustic spectroscopy (QEPAS) and using two quantum cascade lasers as light sources. Single-gas reference spectra of each target mole-cule were acquired and used as PLSR-based algorithm training dataset. The concentrations range explored in the analysis varies from few of part-per-million (ppm) to thousands of ppm. Within this concentration range the influence of the gas matrix on non-radiative relaxation processes can be neglected. Exploiting the ability of PLSR to deal with correlated data, these spectra were used to generate new simulated spectra, i.e. linear combinations of the reference ones. A Gaussian noise distribution was added to the created dataset, simulating the real QEPAS signal fluctuations around the peak value. Compared with standard multilinear re-gression, PLSR predicted gas concentrations with a precision up to 5 times better, even with absorption features with spectral overlap greater than 97%.