Gaussian Process Regression for Quantitative Degree of Polymerization Analysis of Oil-Paper Insulation by EPO-NIRS: A Solution to Field Moisture

In recent years, the near infrared (NIR) spectroscopy has been used in prediction of the degree of polymerization (DP) of oil-paper insulation, but the practices show that the field moisture has a significant influence on the acquired spectra. In this paper, an EPO-GPR model is proposed to eliminate the interference of external moisture and hence to improve the prediction performance of DP of insulating paper in the laboratory as well as on-site experiments. In this model, external parameter orthogonalization (EPO) extracts the differences between the insulating paper samples with varying moisture contents and builds a projection matrix orthogonal to the moisture spectral information; the Gaussian process regression (GPR) with optimized Matérn kernel is employed to mapping the correlation between the spectra and DP of insulating paper. Compared with classical PLS and BPNN models, the EPO-GPR model achieves high prediction accuracy in the laboratory testing (RMSE of 71) and field experiments (RMSE of ∼ 60). The proposed EPO-GPR model can provide a solution to field moisture when using the NIR spectroscopy technique to on-site assess the aging condition of insulating papers.