Deep learning and transfer learning for accurate electrode-level degradation diagnosis in lithium-ion batteries

As a core component of electric vehicles and energy storage systems, lithium-ion batteries require precise performance degradation diagnosis to ensure safe and reliable operation. Traditional state-of-health assessment methods primarily rely on macroscopic indicators such as capacity fade and resistance increase (RI), which fail to reveal complex electrode-level degradation mechanisms. This study proposes a novel deep learning and transfer learning (TL)-based framework for accurate quantification of battery degradation modes (DMs) at the electrode level. The method employs a parallel multi-branch convolutional neural network architecture that simultaneously extracts local and global features from low-rate discharge voltage curves using varying kernel sizes, eliminating the need for manual feature engineering. The model quantifies 4 key degradation mechanisms: lithium inventory loss, anode active material loss, positive electrode active material loss, and RI. Training on 26,521 low-rate discharge voltage curves from lithium nickel manganese cobalt batteries demonstrates high prediction accuracy, with mean absolute errors (MAE) below 2.4% for all DMs, including an MAE of 1.8% for lithium inventory loss. TL further enables model adaptation to lithium nickel cobalt aluminum and lithium iron phosphate battery chemistries, ensuring generalization capability. This work provides an accurate and universal degradation diagnostic tool for battery management systems, facilitating precise health monitoring and predictive maintenance throughout battery lifespan.