Abstract: In recent years, the insulated gate bipolar transistor (IGBT) has been widely used in rail transportation, new energy sources and other fields. Its reliability research is currently a hot topic for scholars. Aiming at the reliability analysis of IGBT, this paper proposes a deep learning model based on long and short-term memory network and convolutional neural network (LSTM-CNN) as the backbone network for IGBT failure prediction. In the model, the introduced attention mechanism gives a higher weight of dominant factors to the features of different dimensions so as to strengthen the influence of important information. At the same time, cross-connection of the network structure fully extracts the features of different levels. The fused multi-level features improve the generalization and robustness of the model. This method is validated on IGBT accelerated aging dataset of National Aeronautics and Space Administration. The experimental results show that, compared with the current mainstream models, the root-mean-square error of the prediction accuracy of the attention mechanism and cross-connection improves by 1.27% and 0.78% respectively. Based on this, a network model based on the fusion of attention mechanism and LSTM-CNN with jump structures is further proposed, and its root-mean-square error of the prediction accuracy increases by 2.68%. It can be concluded that in the failure prediction of IGBT, attention mechanism and cross-connection improve the generalization and robustness of the model from different perspectives, which fully indicates the effectiveness of the proposed method.