Remaining Life Prediction of Bearings Based on Improved IF-SCINet

In the field of health management, predicting the remaining useful life (RUL) of a device becomes critical. However, the RUL prediction process is often affected by a various of confounding factors, resulting in reduced prediction accuracy. To improve the accuracy of RUL prediction, this study first extracts the root mean square, skewness, and Kurtosis from the bearing characteristics, and adopts the multidimensional scale change features to construct a health indicator that fully reflects the bearing degradation trend. Then, a combination of the Isolation forest algorithm and the
$3\sigma $
criterion was used to adaptively determine the first prediction time (FPT) of the bearings. Subsequently, the time series model SCINet is introduced for the first time into the field of bearing life prediction and used to predict the RUL of bearings. Finally, a series of multi-step prediction experiments are conducted on two publicly available datasets, PHM2012 and XJTU-SY, and compared with LSTM, GRU, Informer, and TCN models. The results show that the improved IF-SCINET has a stronger prediction capability compared to the traditional model, which significantly improves the accuracy and stability of bearing RUL prediction.