Threshold alignment indicator driven two-phase nonlinear degradation model for remaining useful life prediction of rolling bearing

Abstract

Wiener process has been widely used in remaining useful life (RUL) prediction studies of rolling bearing by virtue of the ability of quantifying the RUL prediction uncertainty results with the good interpretability and computability. However, the known failure threshold is required for the degeneration modelling of Wiener process. While, the scale drift of health indicator (HI) due to the various degeneration process caused that the preset failure threshold is unavailable, and current wiener-based RUL methods merely focus on the degeneration modelling, especially single-phase, which have seriously affected the RUL prediction results acceptance. To solve this problem, threshold alignment indicator (TAI) driven two-phase nonlinear degradation model (TPNDM) for RUL prediction of rolling bearing is proposed. Specifically, a method named gray relational analysis (GRA) of envelope spectrum singular value (ESSV) is designed to build TAI, in which ESSV is a novel transformed degradation feature space with robustness and trendability to reveal degradation information concealed within the raw vibration data, and GRA is conducted on ESSV to mitigate the scale drift of degradation among different rolling bearings and build TAI. The generated TAI can provide a unitive failure threshold for RUL estimation model and enhance the generalizability of the constructed TPNDM. TPNDM simultaneously considered the factor of nonlinearity, three-variability and two-phase to character the degradation path to enhance the prediction acceptance. Based on the proposed TAI driven TPNDM method, RUL estimation is completed by updating Bayesian criterion. Extensive experiments conducted on both public and industrial scene run-to-failure bearing datasets validated the superiority. These results from comparison experimental show that TPNDM improves MAE, RMSE and Score about 18.09, 21.91, 0.30 individually than some advanced methods, indicating that the proposed TAI driven TPNDM method has more prominent performance for RUL prediction.