Collaborative Prognostics of Lithium-Ion Batteries Using Federated Learning With Dynamic Weighting and Attention Mechanism

The requirement of sufficient run-to-failure data poses a significant challenge in developing deep learning (DL)-based health prognostics for lithium-ion batteries. In practice, numerous battery prognostics datasets are owned by different institutions, and collaboratively leveraging these datasets can provide a promising solution. However, how to utilize prognostics datasets owned by different institutions under data privacy is a challenging task. In this article, a collaborative prognostics method for lithium-ion batteries is proposed, addressing data privacy through federated learning. The approach involves training local models separately based on privately owned prognostics datasets, and these local models are then aggregated to form a global model for collaborative prognostics. To enhance the performance of the collaborative prognostics, a DL-based model equipped with attention free transformer is introduced for battery health prognostics, and a dynamic aggregation method involving weighted averaging of model parameters is proposed to facilitate the model aggregation. To demonstrate the proposed method, aging experiments were conducted on 40 fast-charging lithium-ion batteries in the laboratory. Using this self-tested dataset and two publicly available battery datasets, the proposed method demonstrates satisfactory prognostics results compared to state-of-the-art approaches that do not preserve data privacy.