Remaining useful life prediction (RUL) is one of the key technologies of condition-based maintenance, which is important to maintain the reliability and safety of industrial equipments. While deep learning has achieved great success in RUL prediction, existing methods have difficulties in processing long sequences and extracting information from the sensor and time step aspects. In this paper, we propose Dual Aspect Self-attention based on Transformer (DAST), a novel deep RUL prediction method. DAST consists of two encoders, which work in parallel to simultaneously extract features of different sensors and time steps. Solely based on self-attention, the DAST encoders are more effective in processing long data sequences, and are capable of adaptively learning to focus on more important parts of input. Moreover, the parallel feature extraction design avoids mutual influence of information from two aspects. Experimental results on two real turbofan engine datasets show that our method significantly outperforms state-of-the-art methods.