Degradation prediction of IGBT module based on CNN-LSTM network

Abstract

The reliability of insulated gate bipolar transistor (IGBT) directly affects the safe and stable operation of power electronic system. Based on the IGBT accelerated aging open data set provided by NASA PCoE, the peak collector-emitter voltage during IGBT shutdown is selected as the failure characteristic parameter. Convolutional neural network and long short-term memory network (CNN-LSTM) model is used to predict IGBT failure precursor parameters and estimate the degradation behavior of IGBT modules in this paper. The sliding window method is employed to construct input and output data. When the window width is 3, the prediction model works best. By running the proposed model many times, the average values of MAPE, MAE, MSE and RMSE of the CNN-LSTM network proposed in this paper are 0.0038,0.0468,0.0059,0.0600, which have higher accuracy than other networks. At the same time, the four indicators of the CNN-LSTM model are the most stable and the prediction credibility is higher through the box plot analysis. This prediction method provides a new idea for IGBT degradation behavior prediction.