Remaining useful life prediction of rolling bearing under limited data based on adaptive time-series feature window and multi-step ahead strategy

Predicting the remaining useful life (RUL) of rolling bearings can effectively prevent the breakdown of rotating machinery systems and catastrophic accidents. Most existing RUL prediction methods require massive run-to-failure datasets for modeling. However, it is difficult to obtain these rolling bearing datasets, and for newly or recently deployed rolling bearings, the degradation data are limited with no failure data. Meanwhile, the distribution difference of degradation data of bearings under different working conditions is great, and it is a challenge to employ existing methods to predict RUL. According to the investigation, the health indicators of rolling bearings related to RUL increase exponentially with time. Motivated by this, a novel rolling bearing RUL prediction approach under limited data is proposed in this study. First, a first prediction time (FPT) identification method is developed to obtain the appropriate FPT. Then, the degradation factor is derived mathematically and used to adaptively compress the time-series feature window to better capture the degradation trends of rolling bearings. Subsequently, a stacked bidirectional long short-term memory network (SBiLSTM) is designed to predict and smoothen sequential data. Combined with the degradation factor and SBiLSTM, a multi-step ahead rolling prediction method is presented to predict RUL. Finally, several experiments are conducted on rolling bearings, and the mean absolute percentage error of the proposed method for three representative rolling bearings are 6.77%, 18.92%, and 8.95%, which are superior to other methods. Accordingly, this study contributes to revealing future degradation trends of rolling bearings mathematically, and providing a new idea for implementing other mechanical system prognostics under limited data. • Rolling bearing RUL prediction under limited data is proposed. • Time-series feature window is developed. • Multi-step ahead rolling prediction is presented.