异步通信
计算机科学
国家(计算机科学)
电池(电)
序列(生物学)
限制
估计
荷电状态
实时计算
健康状况
领域(数学)
电池容量
接头(建筑物)
算法
变化(天文学)
人工智能
车辆动力学
估计理论
作者
Shijie Cai,Jia Chen,Jie Hu,Ranjun Huang,Jie Hu,Pei Zhang,Min Wei
出处
期刊:
[Elsevier BV]
日期:2026-11-01
卷期号:: 100490-100490
标识
DOI:10.1016/j.geits.2026.100490
摘要
ABSTRACT Reliable state of charge (SOC) and state of health (SOH) estimation is essential for safe operation and life-cycle management of traction batteries in vehicle-cloud systems. In field data, SOC is inferred at high frequency from operating windows, whereas SOH is updated sparsely from charging events, leading to distinct time scales and observability. Most existing joint-estimation methods assume synchronized state observations or rely on a unified update scale, limiting their ability to capture the persistent influence of low-frequency health states on subsequent charge-state dynamics. This paper proposes SAJE-LLM, a semantic-aware asynchronous joint-estimation framework based on a large language model (LLM), which formulates real-vehicle SOC/SOH estimation as cross-time-scale conditional state estimation. SAJE-LLM builds an event-level asynchronous health memory to propagate historical SOH estimates and health-state embeddings chronologically to later SOC windows. It further develops domain-semantic sequence alignment to map charge-discharge sequences into state-relevant continuous embeddings while preserving local dynamics and cross-feature coupling. A task-state-aware attention adaptation then performs conditioned low-rank correction on key-value projections while keeping the LLM backbone frozen. Experiments on one-year field data from 120 vehicles spanning heterogeneous in-service health states show that SAJE-LLM achieves mean absolute errors of 2.666% and 0.648% for SOC and SOH, respectively, outperforming representative baselines on both tasks. Pack-level validation on a retired battery pack from an unseen vehicle further confirms that SAJE-LLM achieves lower joint-estimation errors under controlled charge-discharge testing.
科研通智能强力驱动
Strongly Powered by AbleSci AI