Remaining Useful Life Prediction of Power Electronic Devices With Physics-Informed Deep Learning and Sparse Data

Accurate remaining useful life (RUL) prediction of silicon carbide (SiC) MOSFETs is essential for ensuring the reliability of power electronic systems, particularly under irradiation environments. However, most existing deep learning approaches rely on densely sampled degradation data, making them unsuitable for sparse-data conditions where degradation observations are limited. To address this limitation, we propose a physics-informed deep learning (PIDL) method designed for sparse RUL prediction. The proposed method integrates total ionizing dose (TID)-induced degradation mechanisms, specifically interface and oxide trapped charge accumulation, into a Transformer-based neural architecture via a customized physicsinformed loss function. This loss explicitly penalizes deviations from on-state resistance degradation trajectories, thereby embedding domain knowledge into the model training process. Subsequently, particle swarm optimization (PSO) is employed to optimize the model hyperparameters. We benchmark our method against a baseline Transformer model without physics-informed components, using four evaluation metrics: mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>), and a composite Score. Under 90% data sparsity conditions, the PIDL approach achieves 27.90% reduction in MAE, 26.51% in RMSE, and 22.90% in Score, demonstrating substantial gains in predictive accuracy and reliability. These results highlight the potential of PIDL in addressing sparse data conditions.