HFTL-KD: A new heterogeneous federated transfer learning approach for degradation trajectory prediction in large-scale decentralized systems

Abstract

Restrictions arising from the limited training data and privacy preservation make large-scale lithium-ion battery degradation trajectory prediction challenging. In this study, a novel heterogeneous federated transfer learning with knowledge distillation approach is proposed for lithium-ion battery lifetime prediction with scarce training data and privacy concerns. The approach enables each device in large-scale decentralized system to not only own its private data, but also a unique network designed based on its resource constraints. Specifically, the central server first designs its unique network according to the resource constraints of each device, and trains the network on publicly available data with entire degradation cycles, thus avoiding the high cost of collecting abundant degradation cycles. Then, the trained model is transferred to each device for collaborative training, in which the knowledge of heterogeneous models extracted by knowledge distillation is used for communication between the isolated devices, rather than the parameters in conventional federated learning. Extensive real-world datasets are leveraged to verify the effectiveness of the proposed approach. The comparison results demonstrate that the proposed method outperforms seven benchmarks. An ablation study indicates that the approach can achieve satisfactory battery residual life prediction while preserving privacy.