Quantitative Analysis of Furfural in Transformer Oil Based on Multilevel FTIR‐Raman Spectral Fusion

This study proposes a multilevel data‐fusion approach combining Fourier transform infrared (FTIR) and Raman spectroscopy for the rapid quantification of furfural in transformer oil. Oil samples representing different aging stages were prepared through accelerated thermal aging of oil–paper insulation, and furfural reference concentrations were determined by high‐performance liquid chromatography (HPLC). Density functional theory (DFT) calculations were further used to interpret the vibrational characteristics of furfural and elucidate the key spectral responses and complementarity between FTIR and Raman spectroscopy. Based on these data, quantitative models were first developed for each spectroscopic modality and subsequently extended to low‐, mid‐, and high‐level fusion strategies. For single‐spectroscopy modeling, the best‐performing FTIR model achieved a prediction‐set coefficient of determination () of 0.9342 and a root mean square error of prediction (RMSEP) of 2.6030 mg/L, whereas the best‐performing Raman model yielded an of 0.9601 and an RMSEP of 2.0253 mg/L. Among the evaluated fusion strategies, mid‐level fusion achieved the best numerical prediction performance. The best‐performing k‐nearest neighbors (KNN) model achieved an of 0.9606 and an RMSEP of 2.0134 mg/L, showing only a marginal improvement compared with the best Raman‐only model ( = 0.9601, RMSEP = 2.0253 mg/L). Within the low‐concentration range of 0–4 mg/L, the root mean square error (RMSE) and mean absolute error (MAE) were 0.6253 and 0.3753 mg/L, respectively, with a bias of −0.0234 mg/L. These results indicate that feature‐band‐based mid‐level FTIR‐Raman fusion provides an effective balance between complementary information integration and redundancy reduction, offering a feasible approach for rapid assessment of transformer insulation condition.