Regenerate large-scale retired second-life battery datasets via recovered capacity labels-based deep learning

Abstract

Accurate diagnosis of the health degradation of retired batteries is crucial for ensuring their safe and reliable reuse. While machine learning offers promising solutions, training models to overcome the high heterogeneity of retired batteries requires massive degradation data, leading to high testing costs and substantial energy waste. Here, we reveal the potential to recover large-scale and high-quality second-life battery datasets from field data to assist in diagnosing the health of retired batteries. By seamlessly fusing deep learning and domain knowledge, we enabled the accurate recovery of capacity labels for operating data without regular fully charging or discharging calibrations. To validate the proposed method, we develop a large-scale degradation test on 96 realistic retired batteries, performing over 50,000 charge/discharge cycles to simulate different stationary energy storage scenarios with 24 charge-discharge intervals. With only 3 capacity measurements available over the second-life, the proposed method accurately recovers the capacity labels with a root mean square error below 30 mAh. Furthermore, the health diagnostic model trained on the regenerated dataset is comparable to the model trained on real data, with an almost negligible error of less than 5 mAh. More importantly, we expect to save at least 98% of test time, electricity, and energy consumption when generating datasets cost-effectively. This study highlights the potential of field data to bridge the critical data gap in diagnosing the health degradation of retired batteries.