A method based on DWRNet and MGAU for RUL prediction of bearing with few samples

Abstract

Accurately predicting the remaining useful life (RUL) of a bearing is crucial in the prognostics and reliability management of machinery. Owing to the cost of wind power operation and maintenance and commercial barriers, collecting early failure samples of gearbox bearings is costly. Accordingly, the prediction of the RUL of bearings with few samples remains a challenging problem. To address this challenge, a two-stage method based on DWRNet and MGAU is proposed to predict the RUL of bearings with few samples. First, a bearing’s health indicator (HI) is constructed using a dynamic weighted residual network (DWRNet), which utilizes a dynamic weighted residual block to fully extract the fault feature of the bearing. Then, a meta-gated adaptive unit (MGAU) neural network is implemented to predict RUL of bearings with few samples via a gated adaptive unit and multi-task learning. Finally, the prediction ability of the proposed method is verified using a dataset of bearings.