Prior task aware-augmented meta learning for early state-of-health estimation of lithium-ion batteries

The accurate estimation of the state of health (SOH) of lithium-ion batteries (LIBs) in their early stages is crucial for the full lifecycle management of electric vehicles and energy storage systems. Current meta-learning methods aim to extract shared information from early degradation data of LIBs to estimate the SOH, which has achieved some success but still faces certain limitations. These methods generally rely on many shared features across tasks; however, they overlook the potential heterogeneity caused by the complex nature of early degradation data, leading to insufficient generalization across different LIBs and consequently affecting the accuracy of early SOH estimation. To address this challenge, we propose the prior task aware-augmented (PTAA) meta-learning model for the early SOH estimation of LIBs. The model captures degradation path features in a task-aware framework and dynamically adjusts the dependencies between features to generate task-specific initialization parameters. Additionally, we design a prior aware-augmented feature network, which consists of an augmented generation module, an auxiliary SOH estimation module, and a prior aware estimation module. This network is optimized through adversarial learning to generate effective prior features, even in the absence of unseen degraded data. Moreover, the augmented generation process is modulated layer-by-layer by a path modulator to ensure that the outputs of the generation module can be quickly adapted to different tasks. Comprehensive experiments are conducted on NASA and CALCE LIBs datasets to validate the effectiveness of the model. The experimental results show that the proposed model improves the RMSE and MAE on the two datasets by more than 48% and 44%, respectively, compared to the existing methods, which validates the superiority and application potential of the model.