The following is an abstract for the selected article. A PDF download of the full text of this article is available here. Members may download full texts at no charge. Non-members may be charged a small fee for certain articles.
Improved Algorithm for Quantitative Analyses of Infrared Spectra of Multicomponent Gas Mixtures with Unknown Compositions
Volume 63, Number 3 (March 2009) Page 338-343
Gianella, Michele; Sigrist, Markus W.
We present a major improvement of an algorithm based on a spectral library search for the quantitative analysis of multicomponent gas samples with unknown compositions. A quantitative spectral database of infrared spectra is used as a training set to compute regression coefficients. Concentrations are computed in the principal component space via principal component regression (PCR). In addition to previous algorithms, we introduce a rating for each candidate substance depending on the concentration found with PCR and a filter that removes candidates that are predicted with negative concentrations if their rating is below a certain threshold. Negative concentrations arise when the measured spectrum contains components that are not contained in the database. The PCR is recomputed until all candidates have a rating above the threshold. Then an adaptive filter "subtracts" the substance with the highest rating from both the measured spectrum and the library and appends it to a hit list. The iteration of this procedure directly produces a list of substances in order of descending importance (i.e., contribution to the measured absorbance) with their corresponding concentrations. The algorithm is tested on spectra of multicomponent surgical smoke samples. The four main components (water, methane, ethane, and ethene) are identified correctly (within the top 5 of the hit list) for an appropriate choice of the rating threshold. The algorithm describes the composition of the smoke sample correctly despite the presence of features in the spectrum that cannot be explained by the spectrum of any single substance present in the database.