Online Real-Time SOH Prediction and Anomaly Detection Under Dynamic Load Conditions and Nonstandard Practice.

For effective battery management, accurate prediction of battery state-of-health (SOH) and timely detection of anomalies is required. Despite recent advances, real-world battery management has yet to overcome such difficult issues as life-cycle SOH prediction and anomaly detection for early warning and fault tracing under realistic operational conditions. In this study, we present an approach for online real-time SOH prediction and anomaly detection for rechargeable batteries throughout their life cycles with a focus on real-world applicability. First, we present a model-based prediction of battery states under normal aging, which serves as a reference for detecting an anomaly. To that end, we propose a method for updating model parameters and their uncertainties cyclically and temporally based on the predicted SOH. In particular, we develop a method for SOH prediction under realistic conditions such as inter- and intracycle variations in load current as well as nonstandard charging and discharging practices. Finally, by fusing the model-predicted state with the measured terminal voltage and current, we achieve a statistically well-defined decision on an anomaly. Experiments using CALCE and custom-collected datasets validate the effectiveness of the proposed method in terms of accuracy and sensitivity for detecting abrupt and slow modes of anomaly.