A Physics-Informed Unsupervised Domain Adaptation Regression Network for Lifetime Prediction of IGBTs

Deep learning (DL)-based remaining useful life (RUL) prediction methods are widely used in the reliability assessment of power semiconductor devices. However, they oftentimes lack the incorporation of physical constraints and laws, leading to less reliable predictions. Moreover, owing to the heterogeneity, DL models achieve accurate lifetime prediction remains challenging in the unknown domain. To address these limitations, taking the insulated gate bipolar transistors (IGBTs) as example, a physics-informed unsupervised domain adaptation regression network (PI-UDARN) is developed for their RUL prediction. In brief, the aging properties of IGBTs are represented from empirical degradation equations and the corresponding loss function constrains the training process of the network. The deep CORAL domain modeling strategy is used to align the various target and source domains. It can also improve the model's generalization and knowledge matching. Experiments on a laboratory dataset to validate the effectiveness of the proposed PI- UDARN.