A Method Based on Time-frequency Disentanglement and LSTM for Remaining Useful Life Prediction of Lithium-Ion Battery

As the technology of new energy vehicles continues to evolve, the health management of lithium batteries within these vehicles is becoming increasingly important. Accurately predicting the remaining useful life (RUL) of lithium-ion batteries (LIBs) is a vital part of this management. This paper presents a method based on time-frequency disentanglement network and LSTM network for accurately predicting the RUL of LIBs with a small amount of labeled data. Initially, a disentanglement network is trained as an upstream model to extract degradation features using momentum contrast learning. Subsequently, these degradation features are fed into a downstream LSTM model to predict the RUL. Finally, filters are employed to smooth the output of the model. We validated our method and conducted extensive ablation experiments using the datasets of three batteries of the NASA’s Prognostics Center of Excellence (PcoE), confirming the effectiveness of our approach and the roles of the disentanglement network and filters in it.