Deep learning driven battery voltage-capacity curve prediction utilizing short-term relaxation voltage

Accurate monitoring of the capacity degradation of batteries is critical to their stable operation. However, evaluating the maximum capacity with limited cycle information alone is insufficient to fully indicate the extent of battery degradation. Here, this study propose a battery degradation monitoring method using relaxation voltage combined with encoder-decoder to extend traditional maximum capacity estimation to the entire voltage-capacity (V-Q) curve estimation. The encoder-decoder is constructed using a two-stage training strategy of unsupervised pre-training and transfer learning. Firstly, the short-time relaxation voltage sequence are input the autoencoder for unsupervised pre-training. Through this auto-encoding process, the encoder acquires feature learning capability on the unlabeled relaxation voltages under the same test conditions. Subsequently, the two-stage training process is completed by freezing the encoder weights and performing transfer learning on the decoder to map the relaxation voltage sequence to its corresponding V-Q curve. The proposed method achieves more advanced prediction performance than direct training at the same epochs. This means higher accuracy in using V-Q curves and the derived incremental capacity curves for comprehensive battery degradation monitoring. Validated on 130 battery samples from different laboratories, the proposed method predicts high-fidelity V-Q curves with a root-mean-square error of less than 0.03 Ah. This study highlights the ability to adopt relaxation voltages for battery degradation monitoring, which is expected to enable fast and comprehensive aging diagnostics in non-constant current charging situations due to the short relaxation time required and without additional cycling information.