Optical spectroscopy is based on the interaction between electromagnetic radiation and matter. As measured spectra of single substances are characteristic like a fingerprint, they can be used to identify atoms and molecules. Concentrations can be determined by the magnitudes of the individual features of the spectra. However, if the investigated sample is composed of different substances, the individual spectra may overlap and are sometimes hard to differentiate. This often applies to larger molecules where the spectra differ only slightly in their characteristic features. In this case multivariate analysis methods can be instrumental to decompose the superimposed spectra. The regression analysis focuses on the correlation between spectral features and concentrations of individual substances. To increase the prediction accuracy of regression, the data set can be prepared by a feature selection or feature projection, which helps to reduce the influence of noise in the data set by reducing its dimension.

We analyzed photoacoustic spectra of mixtures of different volatile organic compounds (VOCs) in the infrared wavelength region. The spectral features of the single substances are broad and overlap strongly. For trace gases with weak absorption, the features of the photoacoustic spectrum are quasi linearly proportional to the concentration, hence linear methods can be applied in this case. Different combinations of feature selection, feature projection and regression are compared to demonstrate their strengths and weaknesses and to determine the combination with the highest detection selectivity.