Ultrasonic feature-based regression of DP, breakdown voltage, and dielectric loss in transformer oil–paper insulation

Oil–paper insulation aging is a major cause of power transformer deterioration, yet key indices, such as paper degree of polymerization (DP), oil breakdown voltage (Ubd), and dielectric loss factor (tanδ), are still mainly obtained through laboratory tests. These indices require individual measurement with distinct equipment, making the whole process tedious and time-consuming. This work investigates whether single-channel ultrasonic measurements of transformer oil can be used to estimate these three indices within a unified regression framework. A controlled thermal-aging experiment was carried out on an oil–paper insulation system at seven aging stages (0–32 days). At each stage, 20 samples were tested, and each sample was paired with its own measured DP, Ubd, and tanδ values. From each waveform, 20 physically interpretable time-, frequency-, and envelope-domain features were extracted. A regularized feature-based regression model was then developed and evaluated primarily by sample-wise stratified fivefold cross-validation, while leave-one-stage-out (LOSO) validation was additionally used as a stricter test of cross-stage generalization. Under the main sample-wise cross-validation protocol, the pooled predictions achieved R2 values of 0.988, 0.827, and 0.991 and NRMSE values of 3.5%, 10.5%, and 3.1% for DP, Ubd, and tanδ, respectively. In the supplementary LOSO evaluation, predictive performance decreased. Overall, these results demonstrate that ultrasonic features enable reliable estimation of multiple insulation indices under controlled laboratory thermal-aging conditions.