Leakage-safe benchmark of machine learning models for IGBT remaining useful life prediction

Reliable remaining useful life (RUL) prediction of insulated gate bipolar transistor (IGBT) power modules is key for condition-based maintenance in industrial drives, but progress is limited by scarce run-to-failure data and overly optimistic evaluation. This paper presents a leakage-safe preprocessing pipeline and a leave-one-device-out (LODO) benchmark on the widely used NASA IGBT accelerated ageing dataset (four devices). Using identical preprocessing and device-aware splits, classical machine learning (ML) models (regression, tree-based ensemble methods) and deep sequence models are benchmarked on a consistent footing. Support vector regression (SVR) delivers the most consistent cross-device performance, with a mean absolute error (MAE) of approximately 0.019 and a coefficient of determination (R2) of 99.3%, while deep learning (DL) models remain competitive but do not consistently outperform the regression baselines under strict leakage-safe evaluation. The results demonstrate that leakage control and device-aware validation are as important as model complexity for achieving reliable generalisation in industrial deployment.