Remaining useful life prediction based on deep residual attention network

Deep learning is gaining growing interests in the field of remaining useful life (RUL) prediction and has achieved state-of-the-art results. Current deep learning-based prognostics approaches, however, do not consider the distinctions of different sensor data during representation learning, which affects their prediction accuracy and limits their generalization. To overcome this weakness, a new deep prognostics network called deep residual attention network (DRAN) is proposed in this paper. DRAN is composed of representation learning sub-network and RUL prediction sub-network. In particular, a new module, i.e., attention module, is constructed in DRAN, aiming to emphasize the important degradation information hidden in sensor data and suppress the useless information during representation learning. The proposed DRAN is validated using the vibration signals acquired by accelerated degradation tests of rolling element bearings. The experimental results show that the proposed DRAN is able to provide accurate RUL prediction results and is superior to some existing convolutional networks.