Knowledge Distillation for State of Health Estimation in Lithium-Ion Batteries

Introduction.

Due to their reliability, long lifespan, and high energy density, Lithium-Ion batteries have evolved to be one of the primary energy storage systems, being utilized in a vast variety of applications. In this context, State of Health (SOH) estimation is crucial for efficient and safe operation of the Battery Management System (BMS), ultimately maximizing battery performance and longevity. Prior work implemented a Bidirectional Long Short-Term Memory (BiLSTM) for reliable SOH estimation across a wide range of Energy Storage Systems (ESS). Within this scope, although Knowledge Distillation (KD) has been used for the compression of SOH estimators and Transfer Learning (TL) has separately improved their multi-domain applicability, the two have not been combined.

Methods.

The framework of this research is founded on the pre-trained, TL-adapted BiLSTM teacher, which is distilled into a compact student model using a combined soft-target and ground-truth loss. The distilled model is then benchmarked against the teacher using RMSE, MAE, and FLOPs/parameter counts on a held-out test set.

Results.

The distilled student significantly reduces computational costs and parameter counts while maintaining SOH error metrics within practical limits, indicating that the TL-adapted teacher’s accuracy can be preserved under substantial compression.

Conclusions.

This work is the first to combine transfer learning and knowledge distillation for SOH estimation. The coupling of cross-condition generalization with compression methods, which together limit the need for extensive condition-specific data, enables the deployment of SOH estimators in hardware-restricted environments for real-time BMS applications without compromising performance.