Inline measuring system for the identification of oil contaminants on metal surfaces using infrared spectroscopy

We present a method for quantitative estimation of oil mixture concentrations on metallic surfaces using reflection-based Fourier transform infrared (FTIR) spectroscopy combined with machine learning analysis methods. Infrared spectra are acquired in reflection geometry on aluminum alloy substrates with root mean square surface roughness values between 0.2 µm and 0.5 µm. Oil mixtures of two representative industrial oils, FERROCOTE 61 A-US and FERROCOTE 61 MAL HCL N2, with varying mixture ratios are applied as thin layers with thicknesses ranging from 0.52 µm to 1.93 µm. Spectra are recorded in the spectral range from 5000 cm-1 to 835 cm-1 using p-polarized light at an angle of incidence of 65 degrees. The resulting absorptance spectra are used as input for a supervised regression model based on partial least squares (PLS). The PLS model is trained on reference samples with known oil mixture ratios and optimized using cross-validation to determine the optimal number of latent components. Once calibrated, the regression algorithm provides a direct mapping from measured FTIR spectra to predicted concentrations of the oil components, enabling quantitative estimation of contaminant composition even for highly similar spectra. Despite geometric variations caused by surface roughness and oil layer thickness, the predicted mixture ratios achieve a root mean square error below 3 % for arbitrary training-test splits using a dataset of 4620 spectra. This methodology may contribute to automated, non-destructive quality control of metal surfaces prior to further processing steps such as coating or joining.