Scholarlink

<p>Remaining Useful Life (RUL) prediction of lithium-ion batteries is fundamental in the assurance of reliability and mission-life in satellite power systems. However, accurate RUL estimation is often challenged by data privacy constraints and the significant variability in battery degradation patterns under different operating conditions, making centralized model training impractical. As a solution to the aforementioned issue, the paper suggests a model of federated learning that does not compromise data privacy and allows joint model training to facilitate strong RUL prediction on heterogeneous battery datasets. The NASA Prognostics Center of Excellence (PCoE) lithium-ion battery aging data, comprising eight 18,650 lithium-ion cells is used to conduct experiments. Multivariate discharge signals such as voltage, current, temperature, time and capacity are converted to a time sequence with a sliding window of 30 cycles. RUL is characterized by an end-of-life capacity threshold and tough chronological data partition is utilized to avoid temporal leakage. To simulate distributed satellite environments, two federated configurations are constructed, each consisting of three client batteries for decentralized training and one unseen battery for global evaluation. A CNN-LSTM-Attention regression model is then fine-tuned on every client and a softmax-based adaptive aggregation method, called as FedWeightedAvg, is offered to mitigate the issue of client heterogeneity by providing greater aggregation weights to the better performing local models. The experimental results show that the proposed method is better in terms of robustness and global prediction error than the standard FedAvg algorithm. The proposed FedWeightedAvg strategy is able to obtain the smaller RMSE equal to 0.1856 in the relatively homogeneous subgroup and 0.6539 in the heterogeneous subgroup, compared to the conventional FedAvg aggregation method. Such findings suggest that adaptive federated aggregation and hybrid deep learning temporal model is an accurate, privacy-preserving, and scalable predictive control over lithium-ion battery RUL in distributed satellite prognostics systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive federated aggregation for robust battery remaining useful life estimation across heterogeneous satellite platforms

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Vankamamidi. S. Naresh,
  • V. S. S. N. Krishna Sai,
  • N. P. S. S. Gayathri,
  • U. L. N. Mahalakshmi,
  • P. B. Vara Lakshmi,
  • B. Sesharathnam

摘要

Remaining Useful Life (RUL) prediction of lithium-ion batteries is fundamental in the assurance of reliability and mission-life in satellite power systems. However, accurate RUL estimation is often challenged by data privacy constraints and the significant variability in battery degradation patterns under different operating conditions, making centralized model training impractical. As a solution to the aforementioned issue, the paper suggests a model of federated learning that does not compromise data privacy and allows joint model training to facilitate strong RUL prediction on heterogeneous battery datasets. The NASA Prognostics Center of Excellence (PCoE) lithium-ion battery aging data, comprising eight 18,650 lithium-ion cells is used to conduct experiments. Multivariate discharge signals such as voltage, current, temperature, time and capacity are converted to a time sequence with a sliding window of 30 cycles. RUL is characterized by an end-of-life capacity threshold and tough chronological data partition is utilized to avoid temporal leakage. To simulate distributed satellite environments, two federated configurations are constructed, each consisting of three client batteries for decentralized training and one unseen battery for global evaluation. A CNN-LSTM-Attention regression model is then fine-tuned on every client and a softmax-based adaptive aggregation method, called as FedWeightedAvg, is offered to mitigate the issue of client heterogeneity by providing greater aggregation weights to the better performing local models. The experimental results show that the proposed method is better in terms of robustness and global prediction error than the standard FedAvg algorithm. The proposed FedWeightedAvg strategy is able to obtain the smaller RMSE equal to 0.1856 in the relatively homogeneous subgroup and 0.6539 in the heterogeneous subgroup, compared to the conventional FedAvg aggregation method. Such findings suggest that adaptive federated aggregation and hybrid deep learning temporal model is an accurate, privacy-preserving, and scalable predictive control over lithium-ion battery RUL in distributed satellite prognostics systems.