Self-attention-based adaptive remaining useful life prediction for IGBT with Monte Carlo dropout

摘要

Insulated gate bipolar transistor (IGBT) is one of the most crucial and fragile components in an electronic system. The remaining useful life (RUL) prediction of IGBTs can precisely forecast the unexpected failure and mitigate the potential risk to guarantee system reliability. In this paper, the IGBTs’ run-to-failure (RtF) aging tests are performed to simulate the degradation process, and a self-attention-based prognostic framework named SA-MCD is proposed for RUL prediction. 21 hand-crafted candidate features are extracted from the transient thermal impedance curve, and half of the sensitive ones are selected to construct the health indicator (HI) based on the self-attention mechanism. Monte Carlo (MC) dropout is combined to provide the confidence intervals by increasing the uncertainty. For standalone online data, the proposed method is proved valid in making high-precision RUL predictions at an early stage. When offline data is available, the adaptation performance is also excellent by updating the model with a small part of initial online data. Through the comparisons with some popular methods, we confirm our proposed method’s superiority.