This study proposes a deep Gamma damage process (DGDP) with rolling Bayesian updating for offline posterior damage-coordinate assessment of rolling bearing degradation. Time-domain, frequency-domain, and time-frequency features are extracted from complete run-to-failure vibration records in the XJTU-SY and PHM2012 bearing datasets. After smoothing, normalization, degradation-direction alignment, and fixed feature screening, nine degradation-sensitive features are retained and a principal-component-analysis-based health indicator (PCA-HI) is constructed. The PCA-HI is then converted into a monotone non-decreasing normalized cumulative damage trajectory, by which the life evolution of each bearing is represented in a posterior damage coordinate from D=0 to D=1. A deep sequence model parameterizes the Gamma distribution of the next damage increment, rolling Bayesian rate updating incorporates observed increments in the offline reconstructed damage prefix, and Monte Carlo threshold crossing assesses the crossing-time distribution on this posterior damage scale. To characterize the applicability boundary and the behavior of the proposed framework under controlled replay, fixed-ratio truncation at 20%, 40%, 60%, and 80%, fixed-initial-window replay, method comparison, and ablation experiments are conducted. The method is intended for offline posterior damage-coordinate assessment and should not be interpreted as calibrated prefix-only online RUL prediction.
Lei et al. (Mon,) studied this question.