A remaining useful life prediction method of SiC MOSFET considering failure threshold uncertainty

Abstract

Different methods have been developed to predict power devices' remaining useful life (RUL). The existing methods need to specify the failure thresholds corresponding to failure precursors of power devices based on historical data. However, there might be heterogeneity in failure threshold between different devices despite being from the same batch, which can severely affect the RUL prediction performance. Aiming at this problem, this article proposes an RUL prediction method based on the non-linear Wiener process considering the failure threshold uncertainty. To incorporate the failure threshold uncertainty into the RUL prediction, the truncated normal distribution is employed to characterize this uncertainty. The maximum-likelihood estimation method is used to estimate the model parameters based on the historical degradation data. Then, the parameters can be dynamically updated by the Bayesian paradigm each time a new piece of condition monitoring (CM) data of the interested device in service is observed. This makes the predicted RUL dependent on the real-time health conditions of the interested device in service. The effectiveness of the proposed method is validated with the power cycling testing results of SiC MOSFETs. Experimental results reveal that the proposed method can reduce the prediction error by 10.273%.