Life evaluation of rolling element bearings using stochastic neighbor embedding deep regression

Abstract

Accurate residual service life (RSL) evaluation of rolling element bearings is significant for prognostics and health management to guarantee rotating machinery safety, availability, and efficiency. This work develops a method called stochastic neighbor embedding deep regression (SNEDR) to enhance the estimation performance of the RSL. First, the appropriate features originating from the vibration data of the tested REB are extracted. The state indicators are subsequently established with the extracted features by introducing the stochastic neighbor embedding. By doing that, the random errors and noises generated from the vibration signals can be minimized, and the evaluation performance may be improved. Finally, the regression model based on the state indicators and the long short-term memory network with time information representation capacity is generated for the RSL evaluation. The availability of the SNEDR is validated by the real data derived from a bear failure experiment. Furthermore, a peer method is introduced for a comparative study. Experimental results show that the SNEDR outperforms the competing method and can yield more reasonable and accurate evaluation results.