Co-Designed Physics-Informed Neural Network FPGA Accelerator for Battery Health Estimation

Physics-informed neural networks (PINNs) provide strong generalization for battery state-of-health (SOH) estimation, but their sine activations make aggressive low-bit quantization and efficient edge deployment difficult. This paper presents a co-designed quantization and FPGA acceleration framework for PINN-based SOH estimation. We introduce per-channel GridWarp quantization-aware training, which learns neuron-wise scale and bias parameters to align the quantization grid with the non-uniform activation distribution. The learned transforms are compiled into per-channel activation LUTs that fuse sine evaluation, GridWarp, and requantization in hardware, while the fully connected layers are mapped to an output-neuron-parallel datapath with DSP packing and deferred offset correction. Across 11 evaluation settings, including six XJTU batches with different charge/discharge protocols, three TJU chemistry batches, and the MIT and HUST datasets, per-channel GridWarp reduces average W4A4 MAPE to 1.93%, compared with 5.51% for post-training quantization, 3.45% for QAT, and 2.41% for per-layer GridWarp. Implemented on a ZYNQ-7020 FPGA, the accelerator delivers 0.69 μs latency at 0.182W, with 37× and 80× lower latency than a desktop CPU and GPU, respectively.

IEEE Consumer Electronics Magazine, 1-8. IEEE