Transferable data-driven capacity estimation for lithium-ion batteries with deep learning: A case study from laboratory to field applications

Abstract

Capacity estimation plays a vital role in ensuring the health and safety management of lithium-ion battery-based electric-drive systems. This research focuses on developing a transferable data-driven framework for accurately estimating the capacity of lithium-ion batteries with the same chemistry but different capacities in field applications. The proposed approach leverages universal information from a laboratory dataset and utilizes a pre-trained network designed for small-capacity batteries with constant-current discharging profiles. By applying this framework, capacity estimation for large-capacity batteries under drive cycles can be efficiently achieved with improved performance. In addition, the incremental capacity analysis is employed on two datasets, selecting a robust voltage interval for health indicator extraction with physical interpretations and uncertainty awareness of different fast charging protocols. The feature extraction and dimension increase processes are automated, utilizing the last short charging sequences in wide voltage intervals while considering the uncertainty related to various user charging habits. Results demonstrate that the proposed strategy significantly enhances both robustness and accuracy. When compared to conventional methods, the proposed method exhibits an average root mean square error improvement of 68.40% and 65.89% in the best and worst cases, respectively. The robustness of the proposed strategy is further verified through 30 randomized health indicator verifications. This research showcases the potential of transferable deep learning in improving capacity estimation by leveraging universal information for field applications. The findings emphasize the importance of sharing knowledge across different capacities of lithium-ion batteries, enabling more effective and accurate capacity estimation techniques.