Edge-Optimized Deep Learning for Real-Time State of Health Estimation of Lithium-Ion Batteries in Embedded Systems

Abstract

Accurate State of Health (SoH) estimation is critical for lithium-ion battery management. While deep learning models like Long Short-Term Memory (LSTM) networks offer superior predictive capabilities, their computational complexity and memory footprints prohibit direct deployment on resource-constrained Battery Management System (BMS) microcontrollers. This thesis addresses this bottleneck by proposing an edge-optimized deep learning framework for real-time SoH estimation. Utilizing features extracted from early-cycle operational data (the first 10 cycles), we develop a Physics-Informed Long Short-Term Memory (PINN-LSTM) architecture as the fullprecision teacher model. To achieve embedded compatibility, we introduce a model compression pipeline integrating Physics-Informed Neural Network Knowledge Distillation (PINN-KD), structural pruning, and W8A32 dynamic quantization, where the knowledge of the PINN-LSTM teacher is transferred to a lightweight LSTM student model. The optimization workflow compiles the models from PyTorch into bare-metal C applications using the Apache TVM stack and Relax Virtual Machine, enabling static memory allocation and Flash-aligned weight storage. Software-inthe- Loop (SIL) on Infineon TC4D7 and Processor-in-the-Loop (PIL) evaluations on NXP MCXN947 microcontrollers validate the system’s real-time performance and concurrent asynchronous data handling. Results demonstrate a massive memory reduction, with Flash storage shrinking from 127.13 KB (full-precision PINN-LSTM teacher) to merely 1.25 KB (quantized LSTM student). Crucially, the W8A32 quantized student model maintains exceptional mathematical fidelity, exhibiting a Mean Absolute Error (MAE) of 3.32×10−8 against the corresponding full-precision LSTM student model. Finally, the study identifies gradient dominance as a fundamental challenge in developing chemistry-agnostic models across mixed LFP, NMC, and NCA datasets, highlighting avenues for future domain-adaptation research.