Remaining useful life (RUL) prediction of bearing plays an important role for rotating machinery operation condition monitoring and it can offer momentous information for machinery system maintenance. The recurrent neural network (RNN) methods have been widely used to predict bearing RUL. However, the RNN based methods cannot address the prediction problem of high-dimension features input data and some critical information might be missed due to its health index development. This paper reports a self-attention sequence-to-sequence network (SASN) to predict bearing RUL, which utilizes the self-attention mechanism and positional encoding to structure the state's relationship of short-term and long-term through the past and current measured high-dimension features. The proposed network consists of an encoder and a fully connected layer. The encoder utilizes a long-time and short-time self-attention mechanism to catch the bearing degradation information, and the fully connected layer maps the embedding from encoder output into the predicted RUL result. The proposed network is suitable for high-dimension features data processing, which can avoid the information loss of artificially health indicator construction of common RUL prediction methods. The proposed method is validated in a case study of the bearing degradation dataset, and the results demonstrate that this method can effectively predict the bearing remaining useful life.