Comparative study of partial least squares and neural network models of near-infrared spectroscopy for aging condition assessment of insulating paper

Abstract

Near-infrared spectroscopy (NIRS) is a rapid and non-destructive detection method for component determination and quantitative analysis with broad applications in numerous fields. In recent years, NIRS has started to be used in the aging condition assessment of power transformers. However, the real applications of NIRS are constrained by the lack of evaluation database and accurate prediction algorithms. In this paper, we aim at comparing different NIRS modeling methods and improving diagnostic accuracy. We build the evaluation database via the preparation of 230 specimens derived from three typical types of insulating paper. Calibration models are established by linear method-partial least squares (PLS) and nonlinear method-back propagation neural network (BPNN) to map the relationship between spectra and the degree of polymerization (DP). The DP prediction results show that using full NIR spectra as the input of the PLS model does not ensure a high prediction accuracy, and it is improved by competitive adaptive reweighted sampling (CARS) that selects the optimal wavelength combinations. Prediction precisions given by BPNN and CARS-BPNN models are shown to be less satisfactory than that of CARS-PLS. We process the original spectra with principal component analysis (PCA) as the input of BPNN and the PCA-BPNN model realizes high prediction precision for three types of paper (RMSE ⩽ 24, r = 0.99). With the identification of paper type by the k-nearest neighbors (KNN) method before prediction, the KNN-PCA-BPNN model solves the problem of the low prediction precision for mixed (unknown) paper samples (RMSE = 36, r = 0.98), which facilitates future field tests as well as related applications in practice.