This technology applies continual learning techniques when secondary battery data from different operating environments and types are input sequentially. It updates the analysis model for new data while retaining information from past data, eliminating the need to retrain existing models from scratch.
Existing deep learning-based battery life prediction models suffer from a sharp decline in performance when operating environments or battery types change after initial training. Retraining the entire dataset to address this increases computational costs and leads to catastrophic forgetting, where previous knowledge is lost.
This technology incorporates continual learning methods such as Elastic Weight Consolidation (EWC), Generative Replay, and Dynamic Expandable Networks. When input battery data—including operating environment, type, or initial defects—differs from existing data or exceeds a data volume threshold, the model undergoes continual learning to preserve past information while ensuring prediction accuracy for new conditions. It can be applied to EV battery diagnostic platforms and used-battery residual value assessment services, reducing the high costs of repeated large-scale retraining whenever new cell models are introduced.
This invention was developed with support from the Ministry of Trade, Industry and Energy for the development of low-cost, high-output VRFB stacks with a high current density of 200mA/cm2 or more.
N/A
