A dynamic incremental learning approach for batteries state of health perception coupled with electrical and mechanical information

Hits:

Impact Factor:7.9

DOI number:10.1016/j.jpowsour.2026.240209

Journal:Journal of Power Sources

Key Words:Lithium-ion battery; State of Health; Mixture of Experts; Transformer; Incremental learning

Abstract:Accurate State of Health (SOH) estimation is vital for lithium-ion battery safety and optimization. Traditional data-driven methods often lack generalization across diverse operational scenarios like slow-charging and fastcharging. This study proposes an incremental learning framework using a Transformer-integrated Mixture of Experts (MoE) architecture. The model integrates multi-source data (expansion and voltage) with lightweight expert networks, enhancing accuracy while reducing memory consumption. For new scenarios, it activates only a sparse subset of experts, enabling efficient knowledge updates without full retraining. Experimental results show the Transformer-MoE model achieves a Mean Absolute Percentage Error (MAPE) of 1.52%, a Root Mean Square Percentage Error (RMSPE) of 2.24%, and a Maximum Absolute Error (MAX-AE) of 6.74%, with only 0.4578 MB memory usage—a 72.31% reduction versus baseline. The proposed incremental learning framework maintains excellent performance on both fast and slow-charging test sets, achieving an overall MAPE of 2.53% and an RMSPE of 3.25%, outperforming traditional data-driven models in most scenarios.

Note:中科院2区

Co-author:Changpeng Tan,Ji Wu

First Author:Li Wang

Indexed by:Journal paper

Correspondence Author:Muyao Wu

Document Code:240209

Discipline:Engineering

Document Type:J

Volume:679

ISSN No.:0378-7753

Translation or Not:no

Date of Publication:2026-04-26

Included Journals:SCI、EI