Pseudo-label augmentation with rated-voltage covariate for Bayesian lifetime estimation of high-voltage insulation from scarce and censored ALT data

초록

High-voltage rotating-machines frequently fail due to insulation degradation; however, accelerated life-test (ALT) data are scarce and heavily censored. Moreover, conventional ALT models analyze each voltage class independently, which limits cross-class extrapolation. To address these challenges, this study develops RIPPLE (Rated-voltage-Informed Pseudo-Probabilistic Lifetime Estimation), a Bayesian accelerated failure-time framework that integrates experimental data and specification information to improve lifetime inference. First, censored specimens are transformed into pseudo-labeled lifetimes by learning a mapping from diagnostic time series and design features to failure time using a lightweight Transformer-MLP model ensemble, with predictive uncertainty explicitly quantified. Second, rated-voltage R is incorporated as a design covariate in a LogNormal rated-voltage-informed Temperature-Nonthermal (R-TNT) stress-life model, enabling joint modeling across heterogeneous voltage classes. Third, pseudo-labels are treated as uncertain soft evidence through a truncated latent-variable likelihood, propagating prediction uncertainty into the posterior and enforcing censoring consistency probabilistically. Applied to multi-stress ALT data on 6.6/13.8 kV form-wound coils, RIPPLE reduced the 90% credible-interval width by approximately 90% at 155 degrees C and 2.1 R relative to the censored-likelihood baseline, while substantially improving point-prediction accuracy (MAPE 0.453 vs. 2.213). Simulation studies further support the effectiveness of the proposed framework.