Explainable Physics-informed Self-attention-based Deep Learning for IGBT Health Estimation

Prognosis strategies have emerged as one of the most popular topics of the reliable framework, enhanced by deep learning in recent years. To address the dilemma of lacking a reliable and explainable deep learning approach for the health monitoring of power devices, we explored the physics-informed deep learning architecture for the health estimation of IGBT, together with an explainable feature weighting mechanism. Combining with the self-attention mechanism, we proposed the physics-informed self-attention (PISA) to assess 4 IGBT’s health state in the case of separating them into training and validation datasets. The result of the experiment has indicated the advances in the effectiveness of the proposed method, which presented a strong generalization capability on the validation data. Furthermore, as the visualization of the internal parameters of the proposed method presented, the capability of self-adapting to the representative features for the health state could be proved by applying the proposed method.