Combined Spectral Evaluation Method for Aging State of Oil-Paper Insulation Based on Channel Attention Improved Variable-Weighted Convolutional Neural Network

The methods on predicting the insulation performance of transformer by measuring the spectra of insulation oil show excellent application value. The aging evaluation method of classification prediction based on single spectroscopy still possesses the potential risk of misjudge, and a more direct and accurate evaluation method of degree of polymerization (DP) value prediction is required. This study proposes a combined spectral method for evaluating the aging state of oil-paper insulation based on a channel attention-variable weighted convolutional neural network (CA-VWCNN). The Fourier transform infrared (FTIR) and Raman spectral datasets of oil-paper insulation were constructed by the accelerated thermal aging experiment at 130 °C, and the two spectral data were fused based on the importance factor block (IF block). The convolution layer extracted spectral variable features, while the channel attention module identified the overall features of the samples to achieve end-to-end prediction of the DP value of insulation paper, and the aging state was evaluated consequently. The results show that the mean absolute error (MAE) of DP predicted value was only 12.66 and $R^{2}$ was 99.09%, which was higher than that of single spectrum. Finally, the decision process of the proposed model was explained and the importance of each spectral variable for DP prediction was visualized. The proposed aging evaluation method has high reliability.