Data-Driven Lifetime Prediction of Power Semiconductor Devices Under EV Mission Profile

Abstract

A key issue during EV traction inverter design is reliable power semiconductor devices, i.e., Insulated Gate Bipolar Transistors (IGBT) and diodes, and lifetime estimation. In a classical method, lifetime estimation is performed through a high-computational, physics-based pipeline including dynamic mission profiles, loss calculation, 4th-order Cauer thermal networks and rainflow counting algorithms. The computationally prohibitive cost per iteration for the traditional technique makes large-scale design-space exploration using such an approach impossible. This paper presents a physics-informed data-driven surrogate model for faster prediction of device lifetime. The primary innovation is the systematic encoding of raw junction temperature waveforms into a 14-dimensional feature vector (statistics and moments, thermal percentiles, dynamic intensity features, fatigue threshold, cycle indicator) which directly maps to device lifetime via a Gaussian Process Regression (GPR) model. The computational burden of the rainflow counting and damage accumulation phases are completely eliminated while still providing the Bayesian uncertainty quantification that the current deterministic sur-rogate techniques cannot. The presented approach is validated via extensive Matlab/Simulink simulations with New York City Cycle (NYCC) and Highway Fuel Economy Test (HWFET) mission profiles. Leave-one-out cross-validation (LOOCV) over 10 heterogeneous missions, reveals a MAPE of 1.94%, MAE of 1.528 years and 153× speedup per-mission in inference runtime (from 2.105 s to 0.0137 s), while the GPR model outperforms all 5 alternative surrogate models for identical missions.