Partial least squares modelling of spectroscopic data from microplasma emissions for determination of CO2 concentration

Abstract

The spectral emissions from a microplasma have been used to predict the CO2 concentration in gas samples covering a concentration range of 0%–100%. Different models based on partial least squares have been evaluated, comparing two different spectral pre-processing filters –multiplicative scatter correction (MSC) and standard normal variate correction (SNV) – and three different wavelength ranges. The models were compared with respect to accuracy, precision, stability and linearity. CO2 samples were mixed with either air or nitrogen. The choice of mixing gas influenced the predicted concentration and basing the models on data from only one mixing gas resulted in higher prediction power. Using air as mixing gas and SNV filtering resulted in a root mean square error of prediction (RMSEP) of 0.03 for an independent test dataset. This RMSEP was of the same range as the experimental error. On the other hand, the models with the best long term stability, reaching the lowest Allan variance, were based on observations with both mixing gases. Models based on MSC filtering generally had slightly higher RMSEP than those based on SNV filtering. Generally, the CO2 concentration could be accurately predicted in the concentration range of 5%–90%. For higher and lower concentrations, the models underestimated the CO2 concentration and were less accurate and precise. Basing the models on fewer wavelengths resulted in reduced linearity. The models were also evaluated by applying them for transcutaneous blood gas monitoring, where they helped to reveal new physiological information.