A Remaining Useful Life Prediction Technique for Rolling Element Bearings Based on Deep Temporal Feature Transfer

This paper presents a bearing remaining useful life (RUL) prediction technique based on an ordered neuron long-short term memory network (ON-LSTM) and a transfer learning. The combined technique can largely improve the RUL prediction accuracy by considering the distribution difference and temporal dependence between the bearing degradation data. In this approach, the ordered feature information within the data is extracted using the ON-LSTM to form the health indicators of the source domain. The distance and distribution of both source and target domains from the bearing degradation data are then adaptively reduced using the transfer learning algorithm. Finally, the degradation trend learned from the source domain can be mapped and transferred to the unlabeled data of the target domain for the bearing remaining useful life prediction. The robustness of the proposed technique is evaluated using the published 2012 IEEE-PHM Challenge bearing lifecycle data. The result shows that the proposed technique can accurately capture the degradation trend of bearings using unlabeled target-domain data.