计算机科学
人工智能
电池(电)
工程类
生成模型
机器学习
生成语法
人工神经网络
钥匙(锁)
集合(抽象数据类型)
作者
Tianjing Wang,Ling Fu,Yuechuan Tao,Zhijun Zhang,Shuyin Lai,Zhao Yang Dong
标识
DOI:10.1109/tte.2026.3696508
摘要
Increased adoption of electric vehicles demands reliable battery health and lifespan predictions for safety and performance. However, current machine learning methods often struggle with limited labeled data and poor generalization across battery types and operating conditions, resulting in training inefficiencies and reduced prediction accuracy. To fill this gap, this study proposes a generative pre-training (GPT) method for battery degradation, named batteryGPT, to enhance the training efficiency and prediction accuracy of tasks for state of health (SOH) and remaining useful life (RUL) estimations. The proposed batteryGPT integrates attention-based multi-modal fusion, dynamic masking, and a Transformer with windowed multi-scale mechanisms during pre-training to capture robust temporal relationships across voltage, current, and temperature data. By leveraging these methods, batteryGPT can learn generalized features from heterogeneous data sources, improving its ability to adapt to diverse operational environments. Fine-tuning adapts the model to specific tasks through task-specific heads, ensuring accurate battery health and lifecycle predictions. Experimental results show that, when fine-tuned with only a few labeled cells, batteryGPT achieves SOH estimation errors below 0.005 and early-cycle RUL prediction errors of about 0.03 across multiple public datasets. These results are obtained under a few-sample fine-tuning setting, while the pre-training stage relies exclusively on unlabeled degradation data.
科研通智能强力驱动
Strongly Powered by AbleSci AI