TY - JOUR
T1 - SOH estimation of lithium-ion batteries subject to partly missing data
T2 - A Kolmogorov–Arnold–Linformer model
AU - Shao, Liyuan
AU - Zhang, Yong
AU - Zheng, Xiujuan
AU - Yang, Rui
AU - Zhou, Wei
N1 - Publisher Copyright:
© 2025
PY - 2025/7/14
Y1 - 2025/7/14
N2 - Accurate estimation of the state of health (SOH) is crucial for improving the safety and reliability of lithium-ion batteries. However, sensor measurements inevitably suffer from incomplete data due to sensor failures caused by factors such as component aging in practical applications. To address this issue, a joint estimation Kolmogorov–Arnold–Linformer (KAL) network model is proposed. Specifically, the Kolmogorov–Arnold Network (KAN) module is employed to replace the Multi-layer Perceptrons (MLP) module in the Linformer model, which enhances the representation of nonlinear features and improves the overall accuracy of the model. A dual-training model approach for SOH estimation is designed, which integrates the health feature (HF)-to-capacity model to infer trends in capacity changes using historical data. Based on these inferred trends, the model is trained to achieve accurate SOH estimation in scenarios with partly missing data. Validation on the publicly available Toyota-MIT-Stanford dataset demonstrates that, compared with other common deep learning methods, the KAL network model exhibits superior accuracy and reliability in scenarios with varying rates of partly missing data.
AB - Accurate estimation of the state of health (SOH) is crucial for improving the safety and reliability of lithium-ion batteries. However, sensor measurements inevitably suffer from incomplete data due to sensor failures caused by factors such as component aging in practical applications. To address this issue, a joint estimation Kolmogorov–Arnold–Linformer (KAL) network model is proposed. Specifically, the Kolmogorov–Arnold Network (KAN) module is employed to replace the Multi-layer Perceptrons (MLP) module in the Linformer model, which enhances the representation of nonlinear features and improves the overall accuracy of the model. A dual-training model approach for SOH estimation is designed, which integrates the health feature (HF)-to-capacity model to infer trends in capacity changes using historical data. Based on these inferred trends, the model is trained to achieve accurate SOH estimation in scenarios with partly missing data. Validation on the publicly available Toyota-MIT-Stanford dataset demonstrates that, compared with other common deep learning methods, the KAL network model exhibits superior accuracy and reliability in scenarios with varying rates of partly missing data.
KW - Kolmogorov–Arnold networks
KW - Linformer
KW - Lithium-ion batteries
KW - Partly missing data
KW - State of health
UR - http://www.scopus.com/inward/record.url?scp=105002302851&partnerID=8YFLogxK
U2 - 10.1016/j.neucom.2025.130181
DO - 10.1016/j.neucom.2025.130181
M3 - Article
AN - SCOPUS:105002302851
SN - 0925-2312
VL - 638
JO - Neurocomputing
JF - Neurocomputing
M1 - 130181
ER -