TD-MSEFormer: temporal dependency multi-scale expert former for IGBT remaining useful life prediction

Accurate remaining useful life (RUL) prediction of insulated gate bipolar transistors (IGBTs) is important for ensuring the reliability of power electronic systems. However, IGBT degradation sequences usually exhibit nonlinear evolution, mixed short-term and long-term dependencies, and stage-dependent degradation characteristics, which makes robust RUL prediction challenging. To address these issues, this paper proposes a Temporal Dependency Multi-Scale Expert Former (TD-MSEFormer). Specifically, a Sequential Dependency Calibration Module (SDCM) is introduced to enhance degradation-related temporal dependency representation. A Multi-Scale Temporal Expert Block (MSTEB) is designed to capture degradation patterns at different temporal scales, while a Temporal Dual-Attention (TDA) block further refines local temporal interactions. Extensive experiments on NASA IGBT Aging Dataset demonstrate that the proposed TD-MSEFormer consistently outperforms other methods, achieving the best overall RUL prediction performance with superior accuracy, robust generalization, and stable degradation representation under complex nonlinear aging conditions.