Pulse-based Continual Learning (PCL) for Second-life Battery SOH Estimation without Catastrophic Forgetting under Data Heterogeneity and Scarcity

Paper i proceeding, 2026

Second-life batteries are critical to sustainable energy transitions, but their diverse historical usage history and limited field data pose major challenges for their reliable health assessment. To address this problem, we propose a pulse-based continual learning (PCL) framework with replay-buffer strategy that preserves prior knowledge while adapting to heterogeneous and data-scarce test conditions. To ensure the rapidness of data curation of second-life batteries, rapid pulse current injection are performed at varying state of charge (SOC) levels, widths, amplitudes, and polarizations. The extracted features from those response voltage signals are utilized as the input for PCL model training. The proposed PCL method treats these pulse current test conditions as domain-incremental tasks and applies a replay-buffer strategy to balance model adaptation and knowledge retention. We performed extensive physical experiments on a broad set of second-life cells, encompassing multiple chemistries (NMC, LMO, LFP), capacities from 2.1 Ah to 35 Ah, and diverse physical forms. The results show that the average SOH estimation accuracy across experimental conditions is over 90%, while the forgetting measure for historical tasks is near-zero. The PCL effectively learns new tasks while retaining established knowledge, delivering accurate and consistent SOH estimates with evolving second-life conditions. Therefore, the proposed PCL method is robust and scalable, providing a pathway toward efficient second-life battery estimation, screening, regrouping with implications for sustainable reusing and recycling.