A Prediction-Driven Framework for Unsupervised Fault Isolation and Diagnostics in LFP Batteries

Predicting lithium-ion battery behavior is essential for optimizing performance and managing degradation. This work focuses on the retrospective detection of regime changes and cell-level fault isolation, validated using historical BMS datasets from two independent LFP battery systems. An LSTM model is trained to predict cell-level voltage, while regime changes are analyzed using unsupervised techniques applied to operating variables to highlight patterns that deviate from nominal operation. The LSTM model achieves low error under stable conditions, but its accuracy declines when system dynamics shift. The localized increase in these prediction residuals effectively isolates the anomalous cell responsible for the deviation, while the unsupervised methods successfully identify the temporal consensus of the regime shift, even without labeled anomalies. Overall, this study provides a practical and scalable approach that combines predictive modeling and unsupervised analysis to isolate temporal regions of interest without needing prior diagnostic information.