Battery life prediction with scarce data using physics-informed data generation and adaptive autoencoder

Abstract

Reliable prediction of battery remaining useful life (RUL) is essential for ensuring the safety and operational efficiency of electric vehicles (EVs). However, current data-driven RUL prediction methods often face significant limitations due to the costly and time-intensive process of collecting complete high-quality degradation data across the battery lifecycle. To address this challenge, a novel framework is proposed, in which physics-informed synthetic data generation is integrated with an adaptive autoencoder-based neural network for efficient RUL prediction. Unlabeled synthetic data reflecting battery degradation behaviors are used to augment the limited training samples. RUL prediction and input reconstruction are jointly performed by the autoencoder to enable structure-aware latent representation learning. Moreover, a dynamically adaptive loss weighting mechanism is introduced to balance predictive accuracy and structural consistency during training. Ablation studies demonstrate that incorporating input reconstruction and synthetic data generation reduces prediction error by over 60% compared to a conventional supervised-only baseline. Furthermore, under data-scarce conditions with only 150 labeled data, the proposed method further outperforms the best semi-supervised baseline by over 15%. These results highlight the robustness and generalization capability of the proposed framework under limited data scenarios.