A Lightweight Foundation Model for Fault Detection of Lithium-Ion Batteries

For lithium-ion batteries, reliable fault detection for charging voltage is essential for operational safety and thermal failure prevention. However, existing battery monitoring solutions face a dual challenge: task-specific models are primarily challenged by limited transferability, while powerful foundation models impose prohibitive computational demands that preclude their integration into resource-constrained edge devices. To address these challenges, this paper proposes a lightweight foundation model for fault detection built upon the IBM Granite TinyTimeMixer (TTM) foundation model. Firstly, we fine-tune the pre-trained TTM backbone with a hybrid loss using only few-shot normal charging sequences, enabling the model to learn the healthy voltage dynamics of batteries. Secondly, a dual-track data pipeline is proposed to adapt to irregular data, where a regular inference grid is generated in parallel with raw asynchronous measurements being retained for preserving vital high-frequency components. Thirdly, a vertical residual alignment mechanism is introduced to align irregular measurements with a continuous prediction curve derived from the TTM model’s grid prediction, enabling precise residual computation despite sampling mismatches. Finally, an empirical 99.99th quantile extreme threshold is calibrated using normal residual distributions to suppress false alarms caused by heavy-tailed sensor noise. Experiments on a lab dataset of 174 battery cells demonstrate that the proposed foundation model detects all fault batteries with zero false positives, which validates its effectiveness and robustness in battery fault detection.