Pattern-Informed Health Indicator Design for Robust RUL Forecasting Via BiGRU Networks

Abstract

PurposeAccurate remaining useful life (RUL) estimation is vital for predictive maintenance of rotating machinery. Existing approaches often rely on predefined degradation trends or fixed health indicators (HIs), which limit their generalizability across diverse operating conditions. This study proposes a robust and interpretable HI formulation directly derived from vibration dynamics, enabling more realistic and consistent degradation modeling for individual bearings.MethodsWe propose binary pattern tracking health indicator (BinTrac-HI), a symbolic-pattern-based framework that encodes vibration segments into binary patterns and extracts three statistical descriptors: dominant pattern ratio, transition density, and skewness. Descriptor weights are optimized via Bayesian hyperparameter tuning guided by the relative accuracy (RA) metric. The resulting HI sequences are exponentially mapped to RUL scores and modeled through bidirectional gated recurrent unit (BiGRU) network for temporal regression.ResultsBinTrac-HI produces smooth, monotonic degradation trajectories across all 11 test bearings in the PRONOSTIA benchmark dataset. When paired with BiGRU, the proposed framework achieves high predictive performance with average metrics of RMSE = 0.053, MAE = 0.045, R-2 = 0.947, RA = 0.798, and Score = 0.863. The method outperforms or rivals state-of-the-art approaches while maintaining lower computational complexity and improved interpretability.ConclusionUnlike previous studies dependent on synthetic HI/RUL assumptions or restricted test cases, this work introduces a scenario-specific and generalizable degradation modeling approach validated across the full PRONOSTIA dataset. The BinTrac-HI + BiGRU framework provides a scalable, interpretable, and noise-resilient solution for practical prognostics in bearing systems.