Battery state of health estimation across electrochemistry and working conditions based on domain adaptation

Accurately assessing the health status of lithium-ion batteries is essential to ensure their safe and efficient application in electric vehicles and energy storage systems. Although various methods have been proposed to achieve battery state of health (SOH) estimation, most of them are only applicable to certain battery types or operating conditions. To address this issue, a novel method is proposed in this study, which leverages data-driven techniques and domain adaptation to cater to different battery electrochemistry and operating conditions. First, the evolution of battery aging is investigated and the incremental capacity sequence with ample aging information is extracted to indicate battery health. Then, the convolutional neural network and bidirectional long-short term memory network are combined to capture the nonlinear relationship between the input sequence and battery SOH. Next, a domain adaptation (DA) based on adversarial training is employed to enhance model adaptability by realizing domain-invariant features extraction. Furthermore, data augmentation is utilized to address data imbalance caused by significant lifespan disparity among different batteries. Finally, datasets with different battery types and aging conditions are used to verify the proposed method. The average estimation error of battery SOH can be controlled within 4.77%, with over 30% reduction contributed by the DA.