PDRUL-Net: a novel physics-informed network of power device remaining useful life prediction

Accurate Remaining Useful Lifetime (RUL) prediction for Insulated-Gate Bipolar Transistors (IGBTs) is critical for power device (PD) reliability and health management, yet challenged by the inherent dilemma between the limited generalization of physics-based models and the poor interpretability/heavy data dependency of data-driven approaches. To bridge this gap, this work proposes PDRUL-Net, a novel physics-informed deep learning framework. PDRUL-Net uniquely integrates the Coffin-Manson-Arrhenius (CMA) degradation model into a deep neural network through a physics-guided architecture featuring parametric embedding and multi-stage modeling (junction temperature estimation, thermal stress characterization, and RUL mapping). A composite physics-informed loss function, incorporating monotonicity and boundary constraints, is specifically designed to explicitly enforce physical laws during training. Evaluated on the NASA IGBT dataset, PDRUL-Net demonstrates significantly enhanced statistical consistency and a substantially higher statistical significance level for RUL predictions compared to baseline models. While achieving competitive predictive performance, maintaining comparable or superior accuracy to state-of-the-art methods, the explicit embedding of physical laws primarily yields improved statistical robustness and interpretability. This study presents an intelligent operation and maintenance solution for PDs, leveraging the synergy of physical credibility and data adaptability for advanced predictive maintenance applications.