Early deployment of deep learning models for lithium-ion battery state-of-health prediction with limited initial data

• The proposed pipeline enables early SOH prediction with limited initial data. • The fine-tuning loop adaptively improves performance as new data become available. • Conv1D–Transformer–LSTM (CTL-Net) is designed for the proposed pipeline. • Fine-tuning achieves 70% reduction in error of the CTL-Net model. Data-driven predictions of the state of health (SOH) in lithium-ion batteries typically require large, high-quality datasets obtained through long-duration cycling experiments. This time-consuming process hinders the early deployment of SOH prediction models in practical applications. To address this challenge, we propose a learning pipeline designed for early model deployment with limited initial data. The pipeline incorporates a real-time fine-tuning loop that progressively improves the prediction performance as new data become available. In addition, we design a hybrid deep learning model, referred to as one-dimensional convolution layer–Transformer–long short-term memory (CTL-Net), to capture both local patterns and long-term nonlinear degradation behaviors. To compensate for limited data availability, a data augmentation technique is employed, enhancing the prediction performance. The initially deployed model achieves sufficient short-term accuracy, but the performance degrades over repeated cycles, and this is addressed through real-time fine-tuning. As additional cycling data become available, the model is fine-tuned in real time, enabling adaptive performance improvements without requiring full retraining. To validate the proposed approach under realistic conditions, initial training is conducted using experimental data from 24 battery cells (18650 format), cycled at 3 C and 20 ± 1°C. Fine-tuning with data from 23 additional cells yields high predictive accuracy ( R 2 = 0.997) and reduces the root-mean-square error by 70% compared with the initially deployed model. Notably, the fine-tuned model outperforms a model trained once on all 47 cells, highlighting the advantage of continuous adaptation. Overall, this study demonstrates that the proposed pipeline enables both earlier model deployment and more accurate SOH prediction in practical battery monitoring scenarios.