Skip to main content
APEC 2026 Main banner

Power Silicon MOSFETs, BJTs, IGBTs, etc.

T12.9 - Uncertainty-Aware Physics-Informed Machine Learning for Remaining Useful Life Prediction of Power Electronic Devices

Wednesday, March 25, 2026
11:40 AM - 12:00 PM CT
Location: 214D
Accurate prediction of the remaining useful life (RUL) of power semiconductor devices is critical for condition monitoring and health management, yet many data-driven models yield overconfident point estimates with poorly calibrated uncertainty. We propose UA-PIDL, a physics-informed, uncertainty-aware Transformer for RUL estimation. UA-PIDL pairs a heteroscedastic Gaussian head to model aleatoric noise with Monte Carlo dropout to quantify epistemic uncertainty. Training follows a two-step scheme: the mean head is first pretrained with mean-squared error, after which variance parameters are learned via negative log-likelihood while the mean remains adaptable, stabilizing learning under limited data. A temporal monotonicity regularizer further encodes the prior that RUL should be non-increasing. Validation on SiC MOSFET power-cycling data, using on-state resistance and threshold voltage as inputs, shows UA-PIDL achieves lower RMSE than strong baselines and delivers calibrated prediction intervals with appropriate coverage, thereby improving the reliability of RUL prognostics.