Machine Learning Based Identification of Lithium-Ion Battery Cell Chemistries

Abstract

With a rapidly growing market for lithium-ion batteries in commercial applications and electrical vehicles, it is foreseeable, that in upcoming years a large number of batteries will return to manufacturers, recycling companies and testing facilities. Unfortunately, recycling companies already struggle with the fast identification of used and unknown batteries. Especially for the material re-use, a profound knowledge of the cell chemistry and its mass fraction is crucial. As of today, this issue is barely discussed in lithium-ion battery research as the focus mainly lies on battery simulation and state estimation.In this work, we present a machine learning based approach for the fast identification of unknown cell chemistries. The approach makes use of differential voltage analysis (dV/dQ) of open circuit voltage (OCV) curves for the training of the machine learning model. The training data was synthetically generated using electrochemical models from PyBaMM. Significant model parameters like the amount of cyclable lithium and the ambient temperature were used to create thousands of differently aged models under different conditions. Besides training the model for lithium iron phosphate (LFP) and lithium nickel manganese cobalt oxides (NMC) cell chemistries, two NMC cells with different mass fractions (811 and 532) were used for training and testing.Different machine learning approaches like decision trees, support vector machine classifiers and different ensemble learning methods were tested and compared against each other. Each of them showed excellent forecast accuracies for the tested chemistries. Moreover, the method was validated by successfully identifying measured OCV curves of real cells. Analysis of the identification techniques showed that the models only need a few significant points to distinguish the chemistries. Hence, the number of OCV points and the duration of the identification time can be drastically reduced for real world applications. In a next step, the OCV curves may even be reduced to partial curves.The proposed method is a promising way of reliably identifying lithium-ion battery cell chemistries in a short time frame. It furthermore provides a straightforward solution for a neglected issue, that will most likely occupy battery research in the future. Figure 1