TF-GIDI: time-frequency granular divergence for interpretable unsupervised grease-lubrication degradation monitoring and relubrication verification in rolling bearings

Grease-lubrication degradation in rolling bearings can develop before localised raceway or rolling-element damage becomes evident, but the associated vibration response is weak, broadband, and often masked by speed- and load-related variations in vibration energy. This study proposes a Time-Frequency Granular Information Divergence Index (TF-GIDI) for unsupervised early monitoring and relubrication-effect verification across operating conditions. TF-GIDI aligns time-frequency representations using condition-specific healthy reference statistics, discretizes them into state fields, and quantifies patch-level distributional shifts relative to a healthy template using Jensen-Shannon divergence. The health indicator provides a scalar degradation trajectory and contribution heatmaps for identifying degradation-sensitive time-frequency regions. In a custom run-to-failure experiment, TF-GIDI detects the start degradation time (SDT) at 540 min, whereas the earliest alarm among eight baseline indicators occurs at 1345 min, yielding an 805 min earlier warning. During relubrication testing, TF-GIDI captures the degradation-recovery-renewed degradation trajectory and quantifies relubrication effectiveness. On public run-to-failure benchmarks, it achieves the lowest remaining useful life (RUL) prediction error. In the time-varying Paderborn LDM case, it reduces mean absolute error (MAE) by 22.8% relative to the second-best baseline indicator. Overall, TF-GIDI supports quantitative non-destructive evaluation (QNDE) by enabling early warning, interpretable degradation localisation, and relubrication-effect verification for grease-lubricated bearings.