Early-stage anomaly detection for lithium-ion batteries via dynamic safety domains and hazard potential integration

Safe operation of lithium-ion batteries under complex loading conditions depends on the ability to recognize incipient anomalies before they develop further. Existing detection methods predominantly rely on instantaneous, hard-threshold judgments of measured quantities, leading to poor diagnostic performance under transient operating conditions with sharp load fluctuations and limited ability to detect early-stage faults. An early-stage anomaly detection framework based on a dynamic safety domain and hazard potential integration is proposed to address this gap. A causal probabilistic network is employed to construct an operating-point-aware safety domain from two marginal conditional distributions, one over the voltage increment and one over the temperature increment. The time-accumulated departure from this domain serves as the alarm criterion. Compared with end-to-end black-box detectors, the framework links each alarm to an explicit sequence of voltage- and temperature-domain departures, which facilitates post-hoc examination of the underlying deviation. A lightweight per-cell affine calibration is further introduced to absorb cell-to-cell dispersion while retaining a single shared network for pack-wide deployment. The framework is validated across diverse battery fault and abuse datasets, demonstrating effectiveness for early-stage faults and abuse events.