Abstract: The infrared spectral data preprocessing methods were coupled with machine learning algorithms to construct multiple combined models for qualitative classification and quantitative regression optimization of various lubricating grease additives. The infrared spectral data of lubricating grease formulation samples designed via orthogonal experimental design were used to train and validate the models, thereby achieving qualitative classification of the types and quantitative regression analysis of the contents of three additives in lubricating grease, namely molybdenum dialkyldithiocarbamate (MoDTC), zinc dialkyldithiophosphate (T202), and potassium borate (T361). The optimal qualitative classification model and quantitative regression model were determined. The results indicate that the combined model of first-derivative (D1) of baseline correction method and random forest (RF) algorithm (D1-RF) achieves the highest classification accuracy for the three additives, with F1 scores of 0.941, 1, and 1 for MoDTC, T202, and T361, respectively, making it the optimal qualitative classification model for lubricating grease additive types. The classification accuracy of the independent modeling system architecture is generally superior to that of the unified modeling system architecture. The partial least squares (PLS) model optimized by genetic algorithm (GA), D1 data preprocessing method, and successive projections algorithm (SPA) (GA-D1-PLS-SPA) yields determination coefficients of fit for the quantitative regression analysis of MoDTC, T202, and T361 contents of 0.8833, 0.9531, and 0.9882, respectively, making it the optimal quantitative regression model for lubricating grease additive contents.

Key words: lubricating grease additives, mid-infrared spectroscopy, qualitative analysis, quantitative analysis, machine learning