Adaptive Extracting Wavelet Transform-Based Damage Indicator-Integrated Improved GPR Model for Bearing RUL Prediction

The bearing is one of the most essential parts of rotating machinery and equipment, and it is of great engineering significance in the accurate prediction of the remaining useful life (RUL) of the bearing. Aiming at this problem, a novel hybrid prediction strategy based on time–frequency representation (TFR) is proposed in this article. First, adaptive extracting wavelet transform (AEWT) is proposed to obtain accurate TFR by creatively exploiting the window width by adjusting center frequency and scale factor based on an attention mechanism and taking the advantage of synchroextracting operator in the IF estimation operator. Then, this article introduces a novel damage indicator (DI) based on time–frequency fractional dispersion entropy (TFDE) to directly extract fault information from TFR to avoid information loss. Finally, the obtained DIs are used as inputs for a prediction model, where the adaptive particle swarm optimization (PSO) algorithm (APSO) is employed to optimize Gaussian process regression (GPR). This method effectively tracks the bearing’s degradation state and accurately predicts its remaining useful life. The experimental verification was carried out with datasets of bearings running to failure. The results show that the proposed time–frequency DI (TFDI) has better monotonicity, and the hybrid prediction model is more accurate in predicting bearing RUL.