Battery prognostics and health management from a machine learning perspective

Transportation electrification is gaining prominence as a significant pathway for reducing emissions and enhancing environmental sustainability. Central to this shift are lithium-ion batteries, which have become the most prevalent energy storage devices. Despite their advantages, the issue of battery degradation during their operational lifetime poses a significant challenge. Progress has been made in modelling and predicting the evolution of nonlinear battery systems using classical physical, electrochemical, first-principle, and atomistic approaches. However, these models are inherently hampered by high computational costs and various sources of uncertainty. Instead of adjusting these classical modelling methods, we argue in this paper for the promising potential of machine learning-based approaches. These approaches allow for the extraction of patterns from inputs and the discovery of complex structures in the target dataset by exploring spatio-temporal features across extensive scales. A significant advancement is the hybrid modelling strategy that merges physical processes with the flexibility offered by deep learning. We present a comprehensive overview of Prognostics and Health Management (PHM) for lithium-ion batteries, with an emphasis on deep neural and kernel-based regression networks. We conclude by offering an outlook on the current limitations, providing a thoughtful analysis of the state of the field and potential future directions.