Combining first prediction time identification and time-series feature window for remaining useful life prediction of rolling bearings with limited data

Limited data are common in the problem of remaining life prediction (RUL) of rolling bearings, and the distribution of degradation data of rolling bearings under different working conditions is quite different, which makes it difficult to predict the RUL of rolling bearings with limited data. To address this issue, this study combines first prediction time identification (FPT) and time-series feature window (TSFW) for predicting the RUL of rolling bearings with limited data. Firstly, the proper first prediction time is identified by a novel FPT identification method considering root mean square and Kurtosis simultaneously. Subsequently, to accurately capture the sequential characteristics of bearing degradation data, the TSFW is constructed and then adaptively compressed considering degradation factor that is derived mathematically. Based on this, this study employs multi-step ahead rolling prediction strategy with degradation factor from FPT to reveal the future degradation trend and then predict the bearing RUL. Finally, the feasibility and generalization of the proposed method under limited data is validated by carrying out several rolling bearing experiments, and the prediction errors for two representative bearings are 14.46% and 8.06%.