A two-phase-based deep neural network for simultaneous health monitoring and prediction of rolling bearings

Simultaneous health monitoring and remaining useful life (RUL) prediction are important objectives in ensuring operational reliability and efficient maintenance of rolling bearings. However, most existing methods ignore the correlation between different degradation stages and RUL, and rarely study the uncertainty quantification of prediction. To overcome these issues, this paper proposes a two-phase-based deep neural network (TPDNN) method, which enables health monitoring and RUL prediction of bearings while providing uncertainty quantification. A logarithmic squared envelope-based diversity entropy is proposed to dynamically evaluate the health status of the bearings, and different degradation stages and RUL labels are adaptively established. Then the feedforward neural network is then used to achieve degradation stage (DS) identification in the first phase. The initial RUL prediction and two kinds of uncertainty quantification are implemented through the bayesian neural network in the second phase. Eventually, the correlation of the DS identification and RUL predictions is handled using a smoothing operator to obtain the final RUL. Experiments and comparisons on two bearing datasets verified that TPDNN has satisfactory prediction performance.