Aiming at the characteristic that the performance of insulated gate bipolar transistor (IGBT) degrades gradually with time under thermal stress, the time series prediction algorithm in deep learning is applied to IGBT fault prediction, a new fault prediction method based on gate recurrent unit (Gru) and principal components analysis-transfer learning (PCA-TL) is proposed. The method takes the Collector Emitter Voltage VCE as the decay parameter, and uses Gru model to construct the mapping relationship between decay parameter and fault time; The PCA technology is used to synthesize IGBT fault indicators with different distribution characteristics, and TL method is introduced. By fine tuning the parameters of Gru prediction model, the migration from source domain to target domain is completed, and the fault prediction of target domain samples is realized. The experimental results show that the fault prediction model based on Gru has higher prediction accuracy and faster training speed than long short term memory (LSTM) algorithm; PCA-TL method can realize the fault monitoring task of similar devices under different working conditions. The feasibility and correctness of the proposed method are verified.