Privacy-preserving federated semi-supervised learning for battery life prediction amid data scarcity

Abstract

Accurate prediction of remaining useful life (RUL) is essential for effective battery management and lifespan optimisation. While recent machine learning approaches offer promising results, their development relies heavily on abundant degradation data with RUL labels, which requires costly run-to-failure tests lasting years. Although massive degradation data are available from millions of batteries in laboratories and in service, access to such data is often restricted due to privacy concerns. Additionally, they usually suffer from quality issues, particularly the absence of RUL labels. To address these issues, we propose a federated-based semi-supervised learning framework enabling collaborative training among diverse battery users that own limited degradation data with RUL labels. This method not only enhances battery RUL prediction by effectively utilising low-cost routine operational data without RUL labels but also protects data privacy across battery users through secure model parameter aggregation. The proposed method is validated on two battery degradation datasets comprising 40 batteries cycled over 24,900 times. Comparative evaluations against federated learning (FL), semi-supervised learning (SSL), and supervised learning (SL) methods are conducted to highlight the effectiveness of our method. Results show that the FL, SSL, and SL methods achieve root mean squared errors (RMSEs) of 27.1, 33.8, and 40.1 cycles, respectively. In contrast, the proposed method achieves an RMSE of 21.3 cycles, resulting in reductions of 21.4 %, 37.0 %, and 46.9 %. This work underscores the potential of federated semi-supervised learning as a practical solution for accurate RUL prediction with reduced battery tests while addressing privacy concerns.