Fault prognostics of lithium-ion batteries in industrial battery management systems

Lithium-ion batteries are critical components in electric vehicles and energy storage systems, where reliable prognostics are essential for ensuring safety and operational efficiency. However, most industrial BMS remain primarily focused on monitoring and protection, with limited capability for early-stage degradation awareness and predictive decision-making. In particular, anomaly detection and RUL prediction are still commonly treated as independent tasks, limiting their potential synergy. To address this limitation, this paper proposes a unified anomaly-aware prognostic framework that transforms degradation deviations into predictive inputs for RUL estimation. A statistically calibrated degradation reference model is first constructed from healthy battery data to characterize nominal aging behavior. Based on this model, a normalized residual-based anomaly representation is defined using confidence-bounded prediction intervals and percentile thresholds. Unlike conventional methods where anomalies are used only for fault indication, the proposed approach reformulates anomaly evolution as an informative signal capturing early degradation dynamics. The resulting anomaly indicators are embedded as exogenous inputs into a MTO-LSTM network to improve learning of temporal degradation dependencies under both normal and abnormal conditions. The framework is validated on Center for Advanced Life Cycle Engineering (CALCE) lithium-ion battery datasets. Experimental results show improved early anomaly detection and higher RUL prediction accuracy compared to standard Long Short-Term Memory (LSTM)-based and hybrid benchmark models, demonstrating the benefit of integrating statistically grounded anomaly evolution into prognostic modeling for health-aware BMS.