Hardware-Accelerated Transformer Framework for Real-Time Battery SoH Estimation

Accurate state-of-health (SoH) estimation is essential for the safety and longevity of electric vehicles and stationary storage. This paper presents a hardware-accelerated SoH framework that deploys a transformer model on a field-programmable gate array (FPGA). While transformers capture long-range dependencies in battery time-series, their computational cost hinders real-time use; we address this with INT8 quantization, deep pipelining, and parallel attention. On a public lithiumtitanate (LTO) dataset (400+ cycles, constant-current discharge at multiple C-rates), the system achieves a mean absolute error of 0.81%, delivers 147 inferences/s, and reduces energy by 13.6 × versus a GPU-based baseline. The model jointly exploits cycle-level capacity and high-frequency voltage signals, enabling early degradation detection with a compact hardware footprint suitable for embedded battery management systems. The approach is chemistry-agnostic and readily extends to lithium-ion and solid-state cells. Results demonstrate the feasibility of real-time, energy-efficient SoH estimation on reconfigurable hardware.