Self-supervised learning for electric vehicle battery remaining useful life prediction using real-world unlabeled data

Accurate and cost-effective prediction of remaining useful life (RUL) is critical for reliable health management of electric vehicle (EV) batteries. However, most data-driven approaches for RUL prediction rely on fully supervised learning and extensive labeling, which is expensive and difficult to scale under heterogeneous operating conditions. This study proposes a contrastive-enhanced variational autoencoder–long short-term memory (VAE–LSTM) framework that leverages large-scale unlabeled charging data in a self-supervised learning paradigm. The framework is pretrained using joint reconstruction and contrastive objectives to learn monotonic degradation representations, and subsequently fine-tuned for RUL regression with limited labeled vehicles. The approach is evaluated on three real-world EV operational datasets, including one heterogeneous fleet dataset comprising passenger cars, taxis, and city buses, as well as two taxi fleets with different scales and operating characteristics. The proposed framework achieves a root-mean-square-error (RMSE) of 27 cycles, outperforming supervised and semi-supervised baselines. Label-efficiency and cross-fleet transfer studies further quantify robustness to domain shift and irregular sampling in field data. With labels from only 30% of vehicles, the pretrained model transfers to two target fleets with RMSE values of 43 and 50 cycles, respectively. A deployment-oriented cost analysis shows that the framework achieves an RMSE within 5% of the fully supervised model while reducing RUL labeling costs by about 70%. Latent factors learned during pretraining are correlated with physically meaningful voltage and energy-throughput signatures, improving interpretability. The proposed VAE–LSTM enables an accurate, interpretable, and economically scalable pathway for real-world EV battery RUL predictions.