A gate-aware GRU model with trend-residual decomposition and quantile regression for remaining useful life prediction of IGBT

Abstract

In predicting the remaining useful life (RUL) of IGBT, the time series data (e.g., emitter-collector on-state voltage drop) characterizing the aging state of IGBT have the nonlinear characteristics of “slow change in the early stage and rapid degradation in the later stage”, and are accompanied by long-term trends and short-term disturbances, which bring challenges to the stability and prediction accuracy of traditional modeling methods. In addition, the existing RUL prediction models mostly focus on point prediction or parameter interval prediction based on distribution assumptions, which is difficult to meet the demand for quantification of prediction uncertainty and risk assessment in engineering applications. To this end, this paper proposes an improved gated recurrent network model (QT-GAGRU) based on trend-residual decomposition and quantile regression for realizing point-interval prediction of IGBT RUL. First, the trend and residual decomposition of the time series are performed and modeled separately using the sliding average method to alleviate the impact of data non-stationarity on the prediction performance; second, the input difference-based gate-aware mechanism (GAGRU) is introduced into the gated recurrent unit to enhance the model's ability of modeling the mutation features in the nonlinear degradation process; and then, for the first time, this paper incorporates the nonparametric interval prediction method -quartile regression (QR) is introduced into the IGBT RUL prediction model to achieve the quantification of model uncertainty without relying on distributional assumptions. Finally, the experimental results on self IGBT aging dataset and NASA public dataset show that the accuracy of the model is better than the existing IGBT RUL prediction models in terms of point prediction, and the interval prediction coverage is high and the interval width is small, which verifies the validity and engineering utility value of the QT-GAGRU model.