Remaining useful life prediction of bearings based on multiple-feature fusion health indicator and weighted temporal convolution network

In prognostic and health management, predicting the remaining useful life (RUL) of bearings is of great significance. The traditional RUL prediction methods have two disadvantages: (a) the health indicator (HI) is mainly constructed through expert experience and signal processing technology, which lacks monotonicity and generalization; and (b) in RUL prediction, the time correlation of information is emphasized, but the influence of the variation amplitude and severity of the vibration signal are ignored. Therefore, a method of bearing RUL prediction based on the multiple-feature fusion HI (MFF-HI) and weighted temporal convolution network (WTCN) is proposed in this paper. MFF-HI is constructed by an MFF depth network (MFFDN) with MISH activation function extracting and fusing the degradation information of bearing time-domain features; WTCN is established based on TCN and a new loss function time-mean square error to reduce the error of network pickup in the bearing degradation stage. The NASA IMS dataset, IEEE PHM 2012 dataset and XJTU-SY dataset are used to verify the superiority of the method. The results show that this method can accurately predict the RUL of bearings with higher prediction accuracy.