Remaining useful life prediction for rolling bearing based on a hybrid data-driven method integrating similarity analysis

Accurately predicting bearings’ remaining useful life (RUL) under severe operating conditions is crucial for predictive maintenance. This work proposes a hybrid data-driven method integrating similarity analysis (HDIS) to enhance interpretability, robustness, and probabilistic RUL estimation. First, deep variational networks are utilized to extract high-fidelity health indicators (HIs) from time–frequency features through unsupervised representation learning. Subsequently, a relevance vector machine processes these HIs to generate sparse representations (SRs), capturing degradation trends while suppressing noise. For RUL prediction, degradation knowledge is transferred from reference bearings using similarity analysis; finally, an empirical degradation model integrates target and reference SRs. This framework enables extrapolation to predefined failure thresholds, yielding high-accuracy probabilistic RUL predictions. Validation on laboratory and benchmark datasets demonstrates the HDIS’s superior robustness and prediction accuracy over single-driven methods, particularly in early degradation phases, while maintaining computational efficiency.