An SVM-Based Health Classifier for Offline Li-Ion Batteries by Using EIS Technology

Monday, 10 October 2022
W. Luo (Cranfied University), A. Syed (Cranfield University), S. Gray, and J. Nicholls (Cranfied University)
With the wide use of electro-vehicles, battery degradation is an unavoidable occurrence, which affects the strategy for the use of battery packs and also the recycling of used batteries. This research presents an offline testing framework and simulation to measure the aging performance of Li-ion batteries within the Battery Management System (BMS) or laddering use for maintenance and recycle activities. It presents the use case of Electrochemical Impedance Spectroscopy (EIS) as a non-destructive inspection method to detect battery state. Multiple cycles (charge and discharge) were performed to acquire EIS curves at different temperatures. The results were captured and digitalised through a suitable circuit model and an automated fitting tool presented in ECS 2018 [1]. The State of Health (SOH) values were calibrated by Coulomb Counting, and data were reshaped as vectors and subsequently used as inputs for a Support Vector Machine (SVM). These data were then used to create a machine learning model which can analyse the aging mechanism of lithium-ion batteries and the decision boundaries are visualised in 2D graphs. The accuracy of these machine learning models can reach 80% in the test cases, and with a good fit to lifetime tracking. The framework allows more reliable SOH estimation in electric vehicles and more efficient maintenance or laddering operations.

Two levels of temperature are adopted, and in each of experiment 12 cells for the same model are packed as two modules. For each module, multiple charge and discharge cycles were performed under the currents of 2.6A (1C) and 5.2A (2C). The EIS results were captured after every 20 charge cycles, with a 20 minute rest, after being fully charged. The aging trend could be observed by observing the changing values of the elements in the Equivalent Circuit Model (ECM) shown below. Therefore, these trends could be summarised and could be regarded as inputs to be analysed by machine learning, combined with their available capacity in each moment. The model shows good generalisation and the ability to summarise different ageing mechanisms for specific models, usage condition and temperatures. This generic approach can also be used for other battery types with different operating conditions to obtain their own unique machine learning models to help determine the ageing of a set of batteries.

This research found that the resistance associated with the solid electrolyte interphase (SEI) layer (R1) and charge transfer resistance (R2) had the greatest effect on the available battery capacity, and that their effect varied at the two temperatures studied. This provides an important basis for machine learning to determine the available capacity at a given temperature. The SVM-based SOH indicator provides a simulation on Li-ion battery aging behaviour at various temperatures, taking all components in the ECM into account. This indicator can be used in applications, such as the laddering of lithium batteries and the initial estimation of SOH, in the battery management system where rapid diagnosis is required. These encouraging simulations represent an initial contribution to the development of offline battery aging monitoring and SOH estimation.

Reference:

[1] M. Murbach, B. Gerwe, N. Dawson-Elli, and L. Tsui, “impedance.py: A Python package for electrochemical impedance analysis,” J. Open Source Softw., vol. 5, no. 52, p. 2349, 2020, doi: 10.21105/joss.02349.