Non-destructive degradation state prediction of IGBT: an explainable framework integrating local variational mode decomposition and differential dynamic transformer

Predicting the degradation state of insulated gate bipolar transistor (IGBT) is of great significance for enhancing the reliability of power electronic systems. Electrical signal based non-destructive testing (NDT) has become an important approach for evaluating its degradation state. However, owing to the complex of IGBT degradation, existing methods are limited by information leakage and insufficient sensitivity to local dynamics, making it challenging to balance predictive reliability and interpretability. Accordingly, a degradation state prediction method integrating local variational mode decomposition (LVMD) and feature-patch differential dynamic transformer (FPDDT) is proposed in this paper. Firstly, the IGBT turn-off peak voltage is subjected to multi-scale decomposition via LVMD. Data leakage introduced by global decomposition is avoided. Secondly, the decomposed mode features are structurally reorganized by FPDDT. The model’s sensitivity to the degradation process is enhanced through differential features. Adaptive fusion of mode features is achieved by combining dynamic convolution. Finally, explainable artificial intelligence (XAI) is used to assess the model’s sensitivity to degradation features and reveal its decision basis. Experimental results show that the proposed method reduces prediction error by at least 42.57% compared with baseline models. Validated the effectiveness of the electrical signal-based NDT method in predicting the degradation state of the IGBT.