Near-infrared spectroscopy (NIRS) is an emerging non-destructive method for quantitatively assessing the degree of polymerization (DP) of transformer insulating paper. To enhance the accuracy and generalization ability of this technique under complex on-site conditions

a deep learning framework based on spectral differential channel concatenation and multi-level attention fusion is proposed. A spectral database consisting of 300 samples with various aging states from six typical paper types is constructed. The first- and second-order differential preprocessing channels are integrated

and data augmentation is performed using baseline drift simulation and random noise injection. A SENet-based module is introduced in the channel domain to adaptively extract the weight of each preprocessing feature

while a Transformer encoder is used in the wavelength domain to capture long-range nonlinear dependencies in the spectral sequences. Results show that the multi-level attention fusion model effectively captures long-range dependencies across spectra and focuses attention on the characteristic absorption regions of C—H and C=O functional groups. The model achieves an RMSE of 45.6 and a MAPE of 5.01% on the test set

outperforming six other benchmark models. In field experiments

the relative error of DP prediction is consistently below 5%

therefore

the proposed method can be used to provide accurate and rapid assessment of insulating paper aging in complex conditions.