Supervised Learning of Synthetic Big Data for Li-Ion Battery Degradation Diagnosis

Models to understand and predict degradation can play a key role in improving the utility of Li-ion batteries. While classical mechanistic models can describe the complex physics of degradation, more recently, data-driven machine learning models have been increasingly utilized for state estimation and lifetime prediction. In this study, the physical grounding of mechanistic models is combined with the power of machine learning via the analysis of published synthetic low rate charge curves generated by a mechanistic model for different thermodynamic degradation modes. The analysis is applied to LFP, NMC, and NCA cells. A step-by-step procedure for developing interpretable machine learning models, including data set splitting, featurization, and model fitting for regression and classification tasks is outlined. Random forest regressors trained on features from incremental capacity analysis of the low rate charge curves can estimate degradation modes with a root mean squared error of 5 %. Further discussion is provided on what can be learned about feature importances, and these leanings are cross-checked with expert defined features.