Remaining useful life prediction of bearings with two-stage LSTM

Bearings are one of the most important rotating machineries of the running parts of high-speed trains, whose failure will cause serious safety problems. Therefore, the prediction of their remaining useful life (RUL) is critical. Stochastic process models and machine learning methods are two commonly used approaches for RUL prediction. Stochastic models, although with high interpretability for degradation mechanism, such as the two-stage degradation features of bearing, show poor ability when the degradation data are polluted with noises from field using condition. On the other hand, machine learning methods, relying on big data and advanced optimization algorithm, can achieve high prediction accuracy. Combining the advantages of both approaches, the paper proposes a two-stage LSTM method that uses statistical feature trends for stage division and LSTM method for RUL prediction. Application on the bearings dataset verified that the two-stage LSTM method is not only more interpretable but also has higher prediction accuracy compared with the traditional LSTM method.