Increasing generalization capability of battery health estimation using continual learning

Accurate and reliable estimation of battery health is crucial for predictive health management. We report a strategy to strengthen the accuracy and generalization of battery health estimation. The model can be initially built based on one battery and then continuously updated using unlabeled data and sparse limited labeled data collected in early stages of testing batteries in different scenarios, satisfying incremental improvement in practical applications. We generate our datasets from 55 commercial pouch and prismatic batteries aged for more than 116,000 cycles under various scenarios. Our model achieves a root mean-square error of 1.312% for the estimation of different dynamic current modes and rates and variable temperature conditions over the entire lifespan using partial charging data. Our model is interpreted by the post hoc strategy with unbiased hidden features, prevents catastrophic forgetting, and supports estimation using data collected in 3 min during ultra-fast charging with errors of less than 2.8%.