计算机科学
学习迁移
人工智能
电池(电)
估计
传输(计算)
工程类
噪音(视频)
机器学习
控制工程
实时计算
数据收集
控制理论(社会学)
作者
Ebubekir Buğra Özarslan,Senem Kurşun
出处
期刊:2017 International Artificial Intelligence and Data Processing Symposium (IDAP)
[Institute of Electrical and Electronics Engineers]
日期:2026-09-01
卷期号:: 1-17
标识
DOI:10.1109/idap71869.2026.11728121
摘要
Accurate State-of-Health (SOH) estimation of lithium-ion batteries is critical for safe, reliable operation of energy-constrained systems including electric vehicles, grid-scale storage, and autonomous platforms. The inherent nonlinearity of electrochemical aging and its strong dependence on chemistry, load profile, and temperature make cross-chemistry generalization a fundamental challenge. Existing transfer learning approaches for battery SOH estimation predominantly focus on cell-to-cell transfer within the same chemistry or rely on full fine-tuning requiring large labeled target datasets. This work proposes a window-based ΔSOH estimation framework that formulates degradation prediction over consecutive Reference Performance Test (RPT) intervals, enabling physically grounded SOH reconstruction without requiring cycle-by-cycle labeled data. A Gated Recurrent Unit (GRU) encoder with a masked Temporal Attention mechanism is trained on two NMC cells (Samsung INR21700-40T, 47 windows), domain-adapted on LFP_CELL_1 (JGNE JGPFR26650), and evaluated on a held-out LFP_CELL_2 under zero-shot and few-shot transfer scenarios. Experiments over five independent random seeds demonstrate that zero-shot transfer achieves a checkpoint-level SOH MAE of 0.373±0.069%, representing an 85.7% improvement over a no-transfer baseline (2.617±0.273%) and outperforming full fine-tuning (0.453±0.207%) in both accuracy and variance stability. Ablation experiments across ten random seeds confirm that the Temporal Attention mechanism reduces checkpoint error by 79.9% relative to attention-free GRU (p < 0.05). Bidirectional transfer analysis (NMC→LFP and LFP→NMC) reveals a 1.43× transfer asymmetry attributed to source dataset richness and OCV feature discriminability. The proposed model comprises only 4,642 parameters (≈18 KB), suitable for embedded BMS deployment.
科研通智能强力驱动
Strongly Powered by AbleSci AI