Abstract: Data-driven methods have achieved some research results in improving the efficiency and accuracy of power system transient stability assessment. However, the transient process of power system involves the change of multi-dimensional time series features. The conventional algorithm has insufficient ability to extract features and lacks interpretability, so it is difficult to reflect the dynamic behavior in the transient process of the system. Therefore, this paper constructs a Transformer model with a two-tower structure. The Transformer encoder is used as a feature extractor. Considering the characteristics of different dimensions at the same time and the influence of each dimension feature on the transient stability of the system at different time steps, it is used as the input of the Transformer model of the two-tower structure to train and learn the influence of each feature channel and time step on the transient stability of the system. Through the fusion mechanism, an end-to-end mapping model from system characteristics to system stability is established, and the high-precision evaluation of transient stability is realized. The decision-making process of the model is explained by attention to heat map visualization. The effectiveness of the proposed method is verified in the IEEE-39 node system.