Artificial Neural Networks and Partial Least Squares Regressions for Rapid Estimation of Mineral Insulating Liquid Properties Based on Infrared Spectroscopic Data

Insulating liquids (transformer oils) are dielectrics used in a wide range of electrical equipment and provide a medium for both insulation and cooling. During equipment operation, liquids are subjected to electrical and thermal stresses. With continued use, they chemically degrade and produce degradation products and aging markers. In this study, models based on Fourier-transform infrared spectroscopic (FTIR) measurements of liquids are proposed for estimating insulating liquid properties (acidity, interfacial tension (IFT), and density) using only a single measurement combined with spectral data analysis. Estimation models based on artificial neural networks (ANN) and partial least squares (PLS) were developed through training and validation on approximately 850 samples of mineral insulating liquids. The proposed models provide an effective means for estimating the acidity, IFT, and density of mineral insulating liquids. The models provide estimation results comparable in reproducibility to standardized laboratory analyses, provide the means for a rapid and accurate assessment of the condition of the insulating liquid, as well as allow the design of dedicated sensors to perform these analyses online.