RUL Prediction of Rolling Bearing via Input Feature Alignment-Combined Different Failure Behavior

Transfer learning methods for remaining useful life (RUL) prediction of rolling bearings have been used for the problem of data distribution discrepancy under variable operating condition and have received sustained attention. However, the data distribution discrepancy still exists between different failure behaviors. To solve this problem, a novel RUL prediction method for rolling bearing under different failure behavior via input feature alignment is proposed. First, an alignment strategy between input features of envelope spectrum is proposed, which makes the data distribution smaller by aligning the input features of the test set and the training set. Meanwhile, it is also considered a normalized way that can be used for online prediction rather than using min-max normalization. Then, a RUL prediction method via the feature alignment strategy-combined failure behavior is proposed, which transforms prediction tasks from variable operating condition to different failure behavior to achieve the better RUL prediction results. Finally, Experiments conducted on IEEE PHM2012 run-to-failure bearing datasets validated the superiority of the proposed method according to the comparative results.