Efficient estimating and clustering Lithium-ion batteries with a deep-learning approach

Abstract Identifying the state of health (SOH) of batteries is critical for estimating batteries from unknown states or grouping retired batteries for secondary utilization to promote recycling sustainability. However, batteries perform distinct capacity fading behaviours in real scenarios and service activities. The energy conversion process is reflected by electrical signals, but proceeds via electrochemical reactions associated with temperature and strain variation. Herein, we develop a deep-learning framework for rapid, accurate, and transferable estimation of battery SOH. This framework incorporates deep neural networks with interconnected electrochemical, mechanical and thermal features as expanded segments. Our model delivers optimal accuracy with mean absolute error (MAE) of 0.519% and root mean square error (RMSE) of 0.795% based on combined features. The model is valid in various conditions and is capable of realizing rapid and accurate prediction merely with one previous cycle. More importantly, the well-trained material domain could be transferred to other electrode systems with few samples. Our work highlights the value of relevant features on battery SOH estimation, and provides guidance on efficient and reliable classification of batteries in the sustainable cycle.