Rapid Fault Detection and Accurate Junction Temperature Estimation in SiC MOSFET-Based EV Drivetrains

The reliability of silicon carbide (SiC) MOSFET-based power converters remains a central challenge in electric vehicle (EV) applications, where fast switching, wide temperature swings, and harsh operating conditions accelerate device degradation and raise the risk of catastrophic failures. This dissertation proposes a unified self-aware framework that integrates ultra-fast short-circuit (SC) protection with real-time junction temperature (Tj) prognostics to improve the transient and lifetime reliability of SiC MOSFETs in EV drivetrains.First, a nanosecond-scale SC detection method is developed by exploiting intrinsic device electrical signatures. The approach monitors rapid variations in drain-source voltage and gate-charge dynamics, enabling the identification of hard-switching fault (HSF) and fault-under-load (FUL) short circuits within 24–150 ns, well below the microsecond-level withstand capability of SiC devices. The circuit is fully integrated into the gate driver with minimal additional components and exhibits strong noise immunity. Experiments on a high-voltage double-pulse setup verify reliable SC detection across wide temperature conditions while reducing voltage overshoot and oscillation duration.Second, a data-driven temperature estimation scheme, UncertTempNet (UTN), is introduced. Multiple temperature-sensitive electrical parameters (TSEPs) are extracted from switching waveforms, and mutual-information-based feature selection identifies the most informative indicators. UTN employs an ensemble neural network with Monte Carlo dropout and attention-based fusion, achieving R² > 99.99% under nominal conditions and maintaining R² ≥ 97.9% under noise and partial-failure scenarios. The uncertainty metric offers built-in diagnostic capability by flagging unreliable predictions.The combined SC protection and Tj prognostics framework is validated on an experimental platform replicating EV-relevant switching environments. Results demonstrate significant improvements in fault resilience, thermal reliability, and self-diagnostic capability, establishing a foundation for future cognitive power electronic systems capable of autonomous health management.