AI-Based Electrochemical Data Reconstruction for Battery Estimation Using On-Board EIS Devices

Development of an On-Board EIS System for Battery State Prediction

One of the most critical aspects of developing a Battery Management System (BMS) for electric vehicles is accurately predicting the battery’s State of Charge (SoC) and State of Health (SoH). However, due to overvoltage effects and hysteresis characteristics during battery charging and discharging, real-time Electrochemical Impedance Spectroscopy (EIS) measurements are difficult to conduct while the vehicle is in motion. To address this issue, this research builds on practical requirements identified while serving as a BMS development advisor for the RACE Hanyang University FSAE team. The goal is to develop an on-board EIS measurement system and establish a protocol for more precise SoC and SoH diagnostics within electric vehicles. Currently, various battery cycling tests and EIS experiments are being conducted to collect data, which serves as the foundation for developing a reinforcement learning-based generative AI model.

Generative AI-Based Electrochemical Data Reconstruction

This research project focuses on designing a Generative Deep Reinforcement Learning model utilizing the Temporal Difference (TD) actor-critic to reconstruct lost electrochemical data. The goal is to restore incomplete Nyquist plots obtained from EIS tests into a more complete form, enabling more accurate battery condition analysis. This approach is expected to enhance the precision of SoC and SoH predictions. This study uses the capacity and impedance data from the aging process of a LiCoO2/graphite 45mAh battery provided by Zhang et al. To validate the practical application of this research, an on-board EIS system is being implemented using Analog Devices' EVAL-AD5941BATZ development board. The ultimate objective is to integrate this system into electric vehicle BMSs, advancing real-time battery state diagnostics technology. Moreover, an integrated BMS system for accurate SOC and SOH estimation was developed by embedding deep learning models into Matlab Simulink, incorporating equivalent circuit model fitting techniques coupled with electrochemical theory.

All photos are taken and edited by me