Silicon carbide (SiC) power modules have become essential components in modern electric vehicle traction systems due to their superior thermal and electrical performance compared to traditional silicon devices. However, ensuring their reliability requires robust condition monitoring to detect degradation before catastrophic in-service failures occur.
Traditional approaches to predicting module failure fall short: physics-based lifetime models cannot operate in real time with sufficient accuracy, while purely data-driven machine learning systems demand large labeled datasets that are expensive to generate and often fail to generalize across different operating conditions and device variants.
A new framework addresses these limitations by merging physics-informed feature engineering with constrained neural networks. Rather than analyzing raw sensor signals, the system transforms measurements into cumulative damage indicators derived from junction temperature fluctuations and a Miner rule accumulator—a well-established fatigue metric from materials science. These features encode the degradation history in an interpretable manner that aligns with how engineers understand device aging.
The framework incorporates two additional refinements. A monotonicity constraint, enforced through gradient penalty regularization, embeds the expected degradation direction as a physics-guided prior, ensuring the model respects physical reality rather than fitting noise. The output uses a heavy-tailed probability distribution rather than a single point estimate, providing calibrated uncertainty quantification that remains robust even when liftoff events or other anomalies introduce non-monotonic perturbations in the voltage signals.
Testing on an industrial power cycling dataset from Infineon Technologies demonstrated significant improvements. The complete framework reduced mean absolute error by approximately 70% relative to conventional data-driven baselines, maintained stability across rigorous cross-validation protocols, and remained lightweight enough for embedded automotive systems. This hybrid physics-machine learning approach offers a practical pathway for deploying real-time health monitoring in next-generation power electronics without requiring the massive datasets or computational resources that purely algorithmic solutions demand.
